1 Introduction

This document describes XRootD configuration directives for the Proxy Storage Service (pss) components.

Configuration directives for each component come from a configuration file. The XRootD structure requires that all components read their directives from the same configuration file. This is the configuration file specified when xrootd was started (see the xrootd -c command line option). This is possible because each component is identified by a unique 3-letter prefix. This allows a common configuration file to be used for the whole system.

The particular components that need to be configured are the file system plug-in (ofs) and the proxy storage plug-in (pss). The relationship between xrootd and the plug-ins is shown on the left. The protocol driver (xrd) runs the xroot protocol which, in turn, utilizes the file system plug-in that, itself, relies on the proxy system plug-in. This is collectively called xrootd, the executable that encapsulates all of the components.

The prefixes documented in this manual are listed in the following table. The all prefix is used in instances where a directive applies to more than one component. Records that do not start with a recognized identifier are ignored. This includes blank record and comment lines (i.e., lines starting with a pound sign, #).

Prefix	Component
ofs	Open File System coordinating acc, cms, oss & pss components
pfc	Proxy File Cache (i.e. specialized proxy plug-in)
pss	Proxy System Service (i.e., specialized oss plug-in)
all	Applies the directive to the above components.

Refer to the manual “Configuration File Syntax” on how to specify and use conditional directives and set variables. These features are indispensable for complex configuration files usually encountered in large installations.

When you are not running a clustered set of proxy servers, you need to specify the ofs.osslib directive. The plug-in is encapsulated in libXrdPss.so and is installed in the standard location. If you are clustering proxy servers (i.e. you specified the all.role directive indicating you are clustering proxies, libXrdPss.so is automatically loaded as the ofs.osslib plug-in.

To properly configure an XRootD proxy you need to review xrd, xrootd, and ofs directives and specify the ones that are relevant to your installation. The xrd and xrootd directives can be found in the “Xrd/XRootd Configuration Reference” while the ofs directives can be found in the “Open File System & Open Storage System Configuration Reference”. If you are configuring a proxy cluster then you will need to specify certain cms directives. These can be found in the “Cluster Management Service Configuration Reference”.

The following section describes the standard proxy service. This proxy service is primarily intended to be used for LAN access to bridge firewalls for remote clients. While it includes a memory caching component that can improve performance for WAN access in limited scenarios, it is not intended to be used as a true WAN proxy. In a subsequent section the disk caching proxy is described. This proxy variant is intended to improve WAN access as it is able to cache files or parts of files on disk at a remote location. Caching files or file segments may minimize WAN traffic depending on the specific workload.

Hence, the standard proxy service is intended to be used as a local proxy while the disk caching proxy is intended to be used as a remote proxy.

A proxy service, whether or not it is a disk caching proxy, can be configured in direct mode or forwarding mode. In direct mode the client is limited to initially contacting a particular endpoint. In forwarding mode, the client is free to specify the actual endpoint that is to be used for the initial contact.

1.1 Direct Mode Proxies

Direct mode proxies are the most common kind of proxies and are suitable to directly fronting an XRootD cluster. The pss.origin directive specifies the initial point of contact. The following diagram illustrates a direct mode proxy.

In the diagram the client present a URL that identifies the proxy server or proxy cluster. The proxy then contacts the data origin, a particular endpoint, on behalf of the client and performs the requested operation. The endpoint can be an actual server or a cluster redirector. This mode is suitable for incoming connections to a fire walled cluster or outgoing connections to a particular data source.

Direct mode proxies can access files (i.e. paths that start with a slash) or data object (i.e. identifiers that do not start with a slash) depending on the specified all.export directives. See the “Xrd/XRootd Configuration Reference” for more information.

1.2 Forwarding Mode Proxies

Forwarding mode proxies are specialized proxies that allow the client to specify the actual endpoint to be used for a request and suitable in cases where the client determines the actual endpoint that must be used. The pss.origin directive specifies that the endpoint will be supplied by the client, as shown below.

In the diagram the client prefixes the destination URL with the proxy location. The proxy server then contacts the specified endpoint on behalf of the client. The client is free to specify any valid endpoint. This mode is most suitable for outgoing connections and poses certain security issues that are described in a subsequent section. A valid destination URL must specify the xroot or root protocol element.

1.2.1 Forwarding File Paths

When configuring a forwarding proxy you should realize that from the proxy’s perspective the URL suffix supplied by the client is initially taken as an actual path. This means that you

a) must export at least some portion of the URL suffix, and

b) can perform authorization operations on the specified URL (see the “Security Configuration Reference” for more information).

Exports of the form

all.export /xroot:/ options

all.export /root:/ options

satisfy the export requirement (see the OFS reference manual for a description of the all.export directive). You can specify longer export names in order to restrict the endpoints that can be contacted using the forwarding proxy. However, since multiple slashes are compressed before comparing the URL against the exported path list, you must not specify successive slashes. For example, say you wanted to restrict forwarding to only the host foobar:1094, then the exports would be specified as

all.export /xroot:/foobar:1094/

all.export /root:/foobar:1094/

Even though the client would actually suffix “/xroot://foobar:1094/” to the proxy’s URL location. The same applies to entries placed in the authorization file. All double slashes must be removed in order for authorization rules to be matched. In both cases, the client’s destination must exactly match the export specification. As a side note, error messages in the proxy’s log file will show URL portions of the path without multiple slashes and should not cause any alarm.

A better alternative is to use the pss.permit directive to restrict the set of target destinations. This allows client to specify the target host in a variety of syntactic ways since the permit applies to the actual resolved destination IP address. In this case you only need to export the protocol segment of the destination URL.

1.2.2 Forwarding Object IDs

To exporting object IDs (i.e. object names without a leading slash) via a forwarding proxy, the configuration file must contain one of the following two export directives:

all.export *

or all.export *?

See the “Xrd/XRootD Configuration Reference” for more information on using object IDs. Since object IDs do not start with a slash; hence, cannot be confused with a file path, the specified client destination URL must appear like an object ID (i.e. it may not start with a slash). Using an absolute file path URL to forward an object ID converts the object ID to a file path and access will likely fail.

Additionally, you can neither use the authorization framework nor the all.export directive to restrict forwarding by destination. Instead you must use the pss.permit directive. This is because object IDs are arbitrary and are only recognized as such in certain restricted contexts.

The following table compares forwarding an absolute path to an Object ID to a server named “foobar:4096” via “myproxy:1094” (notice the missing double slashes for myobject before the second “xroot” and the pathname, myobject).

Access Target	Client URL Specification
/my/data	xroot://myproxy:1094//xroot://foobar:4096//my/data
myobject	xroot://myproxy:1094/xroot://foobar:4096/myobject

A forwarding proxy server is capable of accepting file paths as well as object IDs as long as the correct exports are specified.

1.3 Combination Mode Proxies

A proxy server can be configured in direct more as well as forwarding mode. This is known as a combination mode proxy and is done via a special form of the pss.origin directive. For instance, specifying

pss.orgin = host:port

Allows a client to forward a connection via a URL type of path or connect to a particular destination (i.e. host:port) if the path is not a URL. The exports must allow such combinations. For instance, the trio of

all.export /xroot:/ options

all.export /root:/ options

all.export /atlas options

indicates that xroot: and root: URLs should forward the request to the client-specified destination while paths starting with “/atlas” connect to the destination specified in the pss.origin directive.

In practice, if it is sufficient for the destination host to enforce its exports the whole specification can collapse to a single specification

all.export / options

as the proxy recognizes a URL if it begins with “xrootd:/” or “root:/” and treats anything else as a regular path.

4 Standard Proxy Service

The ofs layer can provide a proxy service by using a special oss plug-in, libXrdPss.so. A proxy service is required if wish to provide public internet access to data served by xrootd servers that can only access a private network or sit behind a firewall. You do not need to setup a proxy service if all of your xrootd servers are accessible via the public network.

As shown in the diagram to the left, a proxy server must have access to the public as well as private or firewalled network. Typically, machines with this capability are called border nodes.

A border node is a computer that either has access to the public as well as the private network using two distinct IP addresses or one that is allowed to access the public network through a firewall using a single local IP address.

The particular scenario in effect is a local option determined by your particular network setup. The biggest challenge is to properly setup the network configuration; a task best left to networking personnel.

Two IP address scenarios require static routes to be established within the border machine so that the local XRootD cluster is always found using the private network while internet traffic uses the public network. Single IP address scenarios (i.e. firewalled networks) require that special firewall rules be established to allow internet access to the proxy server on the XRootD port while disallowing internet access to the local XRootD cluster.

Once routes or firewall rules have been established, you should use the netchk utility, found in the XRootD utils directory, to see if you have access from an external public node all the way to the redirector and each data server node. This verifies that network routing or firewall rules have been correctly established. In order to use netchk, perl must be installed on each machine and you must be able to do an ssh login to each one of them. The utility is described in the last subsection of this section.

The following sections describe how to setup a simple proxy server as well as a proxy server cluster. Additionally, various tuning options are also described. These options are independent of whether you setup a simple proxy or a proxy cluster.

4.1 Simple Proxy Server

The simple proxy server is configured as a non-clustered XRootD server. This means you neither specify the all.role nor the all.manager directives documented in the “Cluster Management Service Configuration Reference”. Other directives may be specified at your discretion.

You control which paths are publically available using the all.export directive that is specific to the proxy server (i.e. the proxy will only allow access to the specified paths). This also means that data servers can provide read-write access to particular paths while the proxy server can only provide read-only access to some or all such paths. The same holds true for the stage attribute. You can prohibit staging files via the proxy by exporting the files via the proxy with the nostage attribute.

The xrd.allow directive can be used to limit which range of hosts can use the proxy server; while the sec.protocol directive can be used to limit access to the proxy server to clients that can only be properly authenticated. These directives do not differ from those you would use in a standard XRootD cluster and documentation can be found the XRootD and security reference manuals.

Very few directives are needed to setup a simple proxy server. A sample configuration file matching the illustrated cluster setup is shown below. Directives shown in red are required.

The ofs.osslib directive tells the ofs layer to use the plug-in that changes a regular XRootD server into a proxy server. Normally, the plug-in is named “libXrdPss.so”. You must specify the location of this shared library plug-in. To support checksum calculations you must also indicate that the checksum manager is a proxy as well. This is done loading the proxy library via the ofs.ckslib directive.

The all.export directive restricts the proxy server to those paths you want to be publically accessible. Most likely, you will also designate these paths as read-only and nostage regardless of how they are locally served within the cluster.

The pss.origin directive tells the plug-in where the XRootD cluster’s redirector is located (in the example it is x.domain.edu). The port number must match the port number used by your local clients (in the example this is 1094). Indeed, the proxy server is, in fact, just another local client that is capable of serving data across the public internet.

The directives, xrd.allow and sec.protocol, are used to limit who can use the proxy server. The xrd.allow directive can restrict access by domain when you use the asterisk notation. The sec.protocol directive can restrict access to authenticated clients. While neither is required, your particular security requirements may force you to specify one or both of them.

4.3.2 Data Caching Directives

pss.cache see memory caching section

pss.cachelib see disk caching section

pss.ccmlib see disk caching section

pss.dca [group | world] [recheck {tm | off}]

Where:

dca [recheck {tm | off}]

Enables direct cache access if the proxy server is configured to be a disk caching server. Direct cache access is only available for clients that can be redirected to an actual file (i.e. release 4.8.0 or above). The parameters are:

group the file mode is set to group read access prior to redirecting the client to directly access the file. This is the default.

world the file mode is set to world read access prior to redirecting the client to directly access the file.

recheck tm The number of seconds between checks whether or not the file is fully cached. The client is redirected to the file on the next read should the file be fully cached. This is only done should the client support redirects on a read request (i.e. release 4.9.0 or above). Specify for tm the number of seconds between checks. You may also suffix tm with h, m, or s to indicate hours, minutes, or seconds, respectively.

recheck off Turns off periodic checking for a fully cached file. This is the default. The client is only redirected to the file upon opening it should it be fully cached

4.3.2.1 Direct Cache Access

The proxy system service, when configured to be a disk caching proxy, supports direct cache access should the file be fully cached. This only makes sense when the cache is directly available to clients using the service. This is normally the case when the cache is located on a distributed file system (e.g. GPFS, Lustre, etc) and all nodes running client applications have the file system mounted at the same mount-point as the proxy service. Direct cache access is a feature of the Proxy Storage Service and must be enabled using the pss.dca directive.

Direct cache access can provide a significant performance enhancement, especially if the file system uses RDMA to transfer data to the client. Even without RDMA, eliminating the server between the client and the data can substantially increase performance.

Direct cache access is only available to clients that support redirection to a file on open (i.e. release 4.8.0). The secondary mode of redirecting the client to a file during read operations is only supported for release 4.9.0 clients. Should a client not support the required redirection level, the Proxy Storage System selectively inhibits redirecting the client and continues to provide the appropriate level of access.

To maintain secure access, clients using direct access should share a group with the user running the proxy cache service. This allows the clients and the proxy service to access the file via a group read access. If this is not possible, you may use the world option, though this would allow anyone to read the file from the cache.

4.3.3 Debugging Directives

pss.debug

pss.trace {all | on | debug}

Where:

debug

Enables debug output in the proxy layer.

trace Enables debug output in the POSIX layer. Each option enables the tracing of events up to a certain level:

all informational events.

on warning events.

debug error events.

4.3.4 Path Processing Directives

pss.localroot path

pss.namelib [-lfncache[src[+]]] [-lfn2pfn] path [parms]

pss.reproxy

Where:

localroot

Configures the default name- to-name mapping using the specified path. The action is identical to that of oss.localroot but is only used by the proxy to map file names prior to requesting action from the origin.

namelib

specifies the name- to-name mapping plug-in library. This directive works just like the oss.namelib directive but only applies to the proxy server. When neither the -lfncache nor -pfn2lfn options are specified, then -pfn2lfn is assumed. Otherwise, you must specify all the required options:

-lfncache Calls the namelib plugin’s pfn2lfn() method to convert the incoming physical filename to a logical one. That filename is used to track the file in any configured cache. This implies that different physical filenames may map onto the same file as long as they have the same logical name.

-lfncachesrc Performs the same action as -lfncache but also passes a single CGI element indicate the data source in URL format as

src=protocol://host[:port]

-lfncachesrc+ Performs the same action as -lfncache but also appends all of the CGI elements that are present on the URL to be used.

-lfn2pfn Calls the namelib plugin’s lfn2pfn() method to convert the incoming logical filename to a physical one. That filename is used to access the file at the specified origin.

reproxy

This directive configures the proxy server for third party copy targeted towards writable origins. This is a very specialized feature and works only with specific storage system (e.g. EOS) and is explained in the next section.

4.3.4.1 Reproxying Third Party Copy

The proxy server is capable of handling third party copy requests (TPC). This allows data to flow through a firewall directly from a source to destination. In most cases this works with no special accommodation for the target endpoint. In some cases, however, certain endpoints require that the proxy be aware of the actual server creating the file. This is required when the proxy is used to also track the progress of the copy. Certain storage system, like EOS, track file creation at the final endpoint and when file status is requested from the cluster redirector, the returned information always shows zero length until the file is closed. Since externally is appears the copy is making no progress the copy operation, in most cases, is terminated.

The mechanism used by the transfer agent to communicate the actual endpoint to the proxy server is:

1) When the transfer agent starts the environmental variable XRD_CPTARGET is set to contain a local path that the proxy server expects to find a symlink that contains the URL of the final endpoint.

2) The transfer agent should create such a symlink once it determines the final endpoint.

3) Whenever an fstat() request is processed for the destination file the proxy server checks for a symlink whose path matches XRD_CPTARGET given to the transfer agent.

4) If the symlink exists, it is read and the symlink deleted. He contents of the symlink is used to issue a stat() request to get the state of the file. The result is returned to fstat() issuer.

5) Should the final endpoint change, the transfer agent should create a new symlink with the correct URL.

Currently, only xrdcp, version 5.3.0 or higher, supports this mechanism.

4.3.5 Networking Directives

pss.inetmode {v4 | v6}

Where:

inetmode

Specifies which network stack to use when connecting to other servers. The default is determined by the -I command line option; which defaults to v6. Options are:

v4 Use the IPV4 network stack to connect to servers. This means only servers reachable by IPV4 addresses may be used.

v6 Use the IPV6 network stack to connect to servers. This means only servers reachable by IPV4 mapped addresses or IPV6 addresses may be used. This is the least restrictive option.

Notes

1) The inetmode option allows you to create network protocol stack bridges. Fr instance, if the proxy accepts IPV6 addresses and inetmode v4 is specified; the proxy becomes a bridge between incoming IPV6 addresses and an IPV4 network.

4.3.7 Tuning Directives

pss.ciosync ssec msec

pss.config {[streams snum] [workers number]}

pss.setopt ConnectTimeout seconds

pss.setopt DataServerConn_ttl seconds

pss.setopt DebugLevel {0 | 1 | 2 | 3}

pss.setopt ParallelEvtLoop number

pss.setopt ParStreamsPerPhyConn number

pss.setopt ReconnectWait seconds

pss.setopt RedirectLimit number

pss.setopt RedirectorConn_ttl seconds

pss.setopt RequestTimeout seconds

pss.setopt TransactionTimeout seconds

pss.setopt WorkerThreads number

Where:

ciosync ssec msec

Configures the parameters for cache I/O synchronization when a file is closed. All outstanding I/O for a file being closed must completed within msec, otherwise the file object is orphaned to prevent the possibility of a crash. The default is 30 seconds for ssec and 180 seconds for msec. This directive is only applicable to a proxy configured with a cache.The parameters are:

ssec The number of seconds between attempts to synchronize outstanding I/O with a file close request. The minimum value is 10 seconds. You may suffix the value with s, m, or h for seconds (the default), minutes, or hours, respectively.

msec The maximum number of seconds that I/O may be outstanding before closure of the file is abandoned and the file object is orphaned. The value must be greater than or equal to ssec times two. You may suffix the value with s, m, or h for seconds (the default), minutes, or hours, respectively

config

Configures various parameters affecting the fuse interface. The defaults are:

streams 512 workers 16

Otherwise, the following option may be specified:

streams The number of parallel streams (i.e. connections) to be used when processing a stat() call. The default is 512. You may specify a maximum of 8192.

workers The number of parallel threads to be used when collecting information from data servers or issuing requests to multiple data servers due to a single request. The default is 16.

setopt

Configures the XRootD client used by the proxy top communicate with the origin. Each setopt directive allows a single configuration argument. Specify one or more of the following, as needed.

ConnectTimeout

The number of seconds to wait for a server connection to complete. The default is 120 seconds.

DataServerConn_ttl

The number of seconds to keep an idle data server socket open. The default is 1200 (i.e. 20 minutes).

DebugLevel

The level of debugging. 0, the default, turns off debug output; increasing values produce more detail.

ParallelEvtLoop

The number of parallel event loops to run to field socket interrupts. The default is 3. Consider increasing this value for servers with a very high transaction rate.

ParStreamsPerPhyConn

The number of parallel TCP streams to use for data server I/O. You can specify a number from 0 to 15. The default is 2.

ReconnectWait

The number of seconds to wait between server reconnects in the presence of a fatal error. The default is 1800.

RedirectLimit

The maximum number of sequential (i.e. without a break) redirects to other servers that may occur before a fatal error is declared. The default is 16.

RedirectorConn_ttl

The number of seconds to keep an idle redirector socket open. The default is 3600 (i.e. 60 minutes).

RequestTimeout

The maximum number of seconds a request can be active without a response before an error is declared. The default is 300 seconds.

TransactionTimeout

The maximum number of seconds of wait-time that can be imposed by the server before an error is declared. The default is 28800 (i.e. 8 hours).

WorkerThreads

The maximum number of responses that can be processed in parallel. Each response is handled in a separate thread. The default is 64.

Notes

1) The proxy also accepts environmental variables that control the underlying XRootD client. Some of these environmental variables offer more extensive control over the client’s behavior then the setopt directive allows. When an environmental variable is set, it over-rides the equivalent specification in the configuration file. The variables are described in the documentation for the xrdcp command.

5 Disk Caching Proxy

The disk caching proxy is almost identical to the standard proxy except that a disk is substituted for a memory cache. This makes the proxy more suitable for WAN access. The libXrdPfc.so shared library is a proxy server cache plug-in used for caching of data into local files. Two modes of operation are supported.

The first mode simply pre-fetches complete files and stores them on local disk. This implementation is suitable for optimization of access latency, especially when reading is not strictly sequential or when it is known in advance that a significant fraction of a file will be read, potentially several times. Of course, once parts of a file are downloaded, access speed is the same as it would be for local XRootD access.

The pre-fetching is initiated by the file open request, unless the file is already available in full. Pre-fetching proceeds sequentially, using a configurable block size (1 MB is the default). Client requests are served as soon as the data becomes available. If a client requests data from parts of the file that have not been pre-fetched yet the proxy puts this request to the beginning of its download queue so as to serve the client with minimal latency. Vector reads are also fully supported. If a file is closed before pre-fetching is complete, further downloading is also stopped. When downloading of the file is complete it could in principle be moved to local storage. Currently, however, there are no provisions in the proxy itself to coordinate this procedure.

A state information file is maintained in parallel with each cached file to store the block size used for the file and a bit-field of blocks that have been committed to disk; this allows for complete cache recovery in case of a forced restart. Information about all file-accesses through the proxy (open & close time, number of bytes read and number of requests) is also put into the state file to provide cache reclamation algorithms with ample details about file usage.

The second mode supports on-demand downloading of individual blocks of a file (block-size can also be passes as cgi information in the file-open request). The distinguishing feature of the second implementation is that it only downloads the requested fixed-size blocks of a file.

While the main motivation is to provide pre-fetching of HDFS blocks (typical size 64 or 128 MB) when they become unavailable at the local site, either permanently or temporarily due to server overload or other transient failures. As HDFS block size is a per-file property, it has to be passed to the proxy on per-file basis as a cgi parameter on the file’s URL Each block is stored as a separate file, post-fixed by block size and its offset in the full file; this facilitates potential reinjection back into HDFS to heal or increase replication of a file-block. Extensions to the HDFS have been developed to allow for an immediate fallback to XRootD access when local HDFS storage fails to provide the requested block.

That said, the mere fact that only referenced blocks are cached makes the caching proxy a powerful tool for any kind of remote file system, irrespective of block size. This is especially true in workloads that reference of a fraction of the file. Because only relevant blocks are cached, disk space usage is minimized and overall WAN performance is improved especially when running multiple jobs that largely reference the same blocks.

When additional file replicas exist in a data-federation, the remote data can be used to supplement local storage, to improve its robustness, and to provide a means for healing of local files. In particular, our intention is to avoid any local file replication of rarely-used, non-custodial data.

Unlike the full-file pre-fetching version, the partial-file proxy does not begin pre-fetching any data until a read request is actually received. At that point a check is made if the blocks required to fulfill the request exist on disk and, if they don't, they get queued for pre-fetching in whole. The client request is served as soon as the data becomes available.

5.1 Configuration

The disk caching proxy is implemented as a specialized caching plug-in that supplants the optional memory caching provided by the standard proxy server.

Disk caching proxy specific directives and options are described below.

pfc.acchistorysize maxcount

pfc.blocksize bytes[k|m]

pss.cachelib path [libopts] required directive

pss.ccmlib path [libopts]

pss.cschk [[no]cache] [[no]net] [off] [[no]tls]

[uvkeep n[d|h|m|s]|lru}]

pfc.decisionlib path [libopts]

pfc.diskusage lowWatermark[k|m|g|t] highWatermark[k|m|g|t]

[files base[k|m|g|t] nom[k|m|g|t] max[k|m|g|t]]

[{purgeinterval | sleep} purgeitvl[h|m|s]]

[purgecoldfiles age{d|h|m|s} period]

pfc.flush {blocks | bytes[k|m|g]}

pfc.hdfsmode [hdfsbsize bytes[k|m]]

pfc.osslib path [libopts]

pfc.prefetch numPrefetchingBlocksPerFile

pfc.ram bytes[m|g]

pfc.spaces data metadata

pfc.trace {none | error | warning | info | debug | dump}

pfc.user username

pfc.writequeue maxblks nthreads

Required Directive

cachelib path [libopts]

specifies the path of he shared library than contains the disk caching proxy plug-in. It may be followed by options, if any. Currently, there are no options.

Optional Directives

pfc.acchistorysize maxcount

Specifies the number of access records kept for each data file. Specify for maxcount a value between 20 and 200, inclusive. The default is 20.

blocksize bytes[k|m]

sets the block size used by proxy. Specify for bytes the size of a block. The quantity may be suffixed by k or m to indicate kilo- or megabytes, respectively. All read requests, including prefetching are rounded up into requests of this size and aligned to block boundaries. Read requests that cross one or more block boundaries are issued as separate read requests. This is also the size that gets written to disk in a single operation. The default is 128k.

Be aware that when client requests are processed using asynchronous I/O, the incoming request is broken into segment sized pieces to allow the data to be streamed from the origin. The default segment size is 64k. The only adverse effect of this mismatch is that the hit/miss statistics for a particular read will not match the expected result. Setting the xrootd segment size to match the default block size will degrade the transfer rate, especially if the origin is behind a firewall. Consider adjusting the block size to match the segment size if you need to eliminate the hit/miss aberration.

ccmlib path [libopts]

specifies the location of the shared library that contains the cache context manager plugin provides custom cache semantics. Specify for path the location of the ccm plugin shared library and, optionally, plugin specific options for libopts.

cschk { [[no]cache] [[no]net] [off] [[no]tls] [uvkeep {n[d|h|m|s]} | lru}}

specifies data integrity options. The cache has the capability of verifying that the correct data was read from the network (data in transit integrity) as well as the correct data was read from the permanent cache (data at rest integrity). Refer to the following section on a detailed explanation of these options. The options are:

[no]cache do [not] perform cache checksum integrity checking; nocache is the default.

[no]net do [not] perform network checksum integrity checking; net is the default.

[no]tls do [not] use TLS as a substitute for network checksum integrity checking when a server supports TLS but does not support data checksums; tls is the default.

off synonym for the combination nocache nonet.

uvkeep specifies how long cached files with unverified checksums are to be kept. Specify for n the number of days, hour, minutes, or seconds (the default) after which such cached data is to be deleted. A value of 0 never records unverified checksums and their associated data to permanent media (i.e. such data is served only via a memory cache). Specify lru, the default, to use the standard mechanism for deleting cache data files.

decisionlib path [libopts]

specifies the location of the shared library that contains the cache decision plugin which determines whether or not to cache a file. All files are cached if a plugin is not specified. Specify for path the location of the decision plugin shared library and, optionally, plugin specific options for libopts.

diskusage

specifies physical storage management properties. The parameters and options are (see the notes for additional caveats):

lowWatermark[k|m|g|t] highWatermark[k|m|g|t]

specifies the total usage on the file system or set of disks configured with the oss.space directive that triggers purging of cached files (highWatermark) and the total usage at which the purging is stopped (lowWatermark). The watermarks must be specified as a decimal fraction of total available space (default values are 0.90 and 0.95) unless the values are suffixed by k, m, g, or t. In which case, they must be absolute sizes in k (kilo-), m (mega-), g (giga-), or t (tera-) bytes, respectively.

files base[k|m|g|t] nom[k|m|g|t] max[k|m|g|t]

Specifies limits for actual usage of disk space by the data files owned by the configured cache instance. When files reach max, they are purged down to nom. If purging due to total disk usage requires further removal of files, they will be purged down to base if required for reaching of lowWatermark disk usage. Specify absolute sizes suffixed by k (kilo-), m (mega-), g (giga-), or t (tera-) bytes. Default: limits are not set and file usage limits are not checked nor enforced.

{purgeinterval | sleep} purgeitvl[h|m|s]

specifies the interval between subsequent disk usage checks. Specify for purgeitvl a number optionally suffixed by h for hours, m for minutes, or s for seconds. The resulting value must be between 60 and 3600 seconds. The default is 300s. The keyword sleep is a deprecated synonym.

purgecoldfiles age{d|h|m|s} period

specifies the time since last access at which a file is unconditionally purged. This is useful for cache instances that share disk space and are only used sporadically so it makes sense to clear the cache between peak usages. Specify for age a number optionally suffixed by h for hours, m for minutes, or s for seconds. The resulting value must be between 1h and 360d, inclusive. The period argument specifies how many purge cycles need to occur before an age check if performed. This avoids redundant scanning of meta-data when other purge conditions (disk and file-usage based ones) are not met. There is no default and cold files are not purged unless this parameter is specified.

flush {blocks | bytes[k|m|g]}

forces a sync operation on all of the data and meta-data files after the number of specified blocks or bytes have been written to memory. The bytes value may be suffixed by k, m, or m to indicate kilo-, mega-, or gigabytes, respectively. The default is 2000 blocks.

hdfsmode [hdfsbsize bytes [k|m]]

enables storage of file fragments as separate files. This is specifically used for storage healing in HDFS. Specifying bytes sets the default fragment size. Specify for bytes the size of a fragment. The quantity may be suffixed by k or m to indicate kilo- or megabytes, respectively. The default is 128MB.

osslib path [libopts]

specifies the location of the shared library that contains the storage system plugin that should be used for all cache operations (i.e. read, write, etc). Specify for path the shared library that contains the plugin and, optionally, plugin specific parameters for libopts. The default is to use standard oss plug-in for all file system functions. Refer to the “Open File System & Open Storage System Configuration Reference” on how to configure the standard oss plug-in.

prefetch num

maximum number of pre-fetching blocks per file. The default is 10. Set this value to 0 to disable pre-fetching.

ram bytes[m|g]

maximum allowed RAM usage for caching proxy buffers that still need to be written to disk. Beyond that point the cache will serve further read requests by forwarding them to the remote server (note, it still needs to allocate buffers for this). For clients, specify a value between 256m and 64g (default is 256m). For servers, specify a value between 1g and 256g (default is 1g).

spaces data metadata

spacifies the names of oss space name for data and the oss space name for metadata. The oss.space directive is used to create spaces. Default value of both is public.

trace {none | error | warning | info | debug | dump}

Set severity level of caching proxy log messages. The levels are listed in increasing verboseness. The default value is warning.

user username

specifies the user name to pass to the osslib plug-in when accessing the disk cache. Specify a valid Unix user name for username. Normally, this should be the user name of the proxy daemon.

writequeue maxblks nthreads

specifies how disk cache write-backs are to be handled. Specify for maxblks the maximum number of blocks to be taken off the pending write queue and written to disk in a single iteration of the disk-writer loop. Accepted values are between 1 and 1024. The default is 16 for a regular proxy server and 8 for a server-less proxy cache. Specify for nthreads the number of threads that perform writing of cache blocks to disk. Accepted values are between 1 and 64. The default is 4 for a regular proxy server and 1 for a server-less proxy cache.

Notes

1) All features of the standard Open Storage System are available to the disk caching proxy. Perhaps the most essential feature is logical volume management that allows the concatenation of disk partitions into a single logical volume. Thus, you can easily extend the disk cache. See the oss.space directive in the “Open Storage System Configuration Reference”. There are many other oss options that may also be helpful in extending the features of the disk caching proxy.

2) If several caching servers are running on the same file system or set of disks configured with oss.space directive, they should all use the same lowWatermark, highWatermark, and purgeinterval. They can use different settings for the files section and can even oversubscribe the disk usage for nom usage as the low/highWatermark mechanism will kick in to ensure specified amount of disk remains free. The baseline values for all the instances should add up to less than lowWatermark.

3) The tls option is ignored unless net is in effect.

4) The options are cumulative across directive specifications.

Examples

a) Enable file prefetching:

pss.cachelib libXrdPfc.so

pfc.ram 100g

pfc.blocksize 512k

pfc.prefetch 8

b) enable hdfs healing mode, with block size 64 MB:

pss.cachelib libXrdPfc.so

pfc.ram 100g

pfc.hdfsmode hdfsbsize 64m

5.1.1 Disk Caching S3-type Objects

The disk caching feature can be used to also cache S3-type data objects (e.g. Ceph objects). Such objects are identified by object ID and do not start with a slash. You must specify certain configuration options to enable successful caching of S3-type objects. Minimally, you need the following set of directives in addition to any other applicable directives.

all.export *?

pss.namelib -lfncache libXrdN2No2p.so [parms]

parms: [-maxfnlen len] [-slash {c | xx}] [prefix]

The all.export directive specifies that object ID’s are allowed and that any CGI information (i.e. characters after the first question mark) are to be removed from the object ID and passed as separate information to the proxy plugin. The plugin requires that any CGI information be passed separately so must specify ‘*?’ to get consistent results.

A special Name2Name plugin must be used to convert object ID’s to file paths so that the data they refer to can be cached as local disk files. You don’t need this plugin if the storage system plugin you specified using the pfc.osslib directive is an object store (e.g. Ceph). The default cache storage system is a file system store via the oss plugin.

The plugin handles object ID conversion as follows:

1) All occurrences of a slash (i.e. ‘/’) are changed to be a reverse slash (i.e. ‘\’ or escape). This is necessary to prevent object id’s containing slashes from colliding. If your object ID’s use reverse slashes then you must choose another character that you won’t use in the ID’s and specify that character using the ‘-slash’ option in the parms passed to the plugin. You can specify it as a single character or as a two character hexadecimal code (e.g. 32 for a space character).

2) The maximum allowed filename length is determined by what is allowed to be created in ‘/’ directory. If this is incorrect, use the -maxfnlen parms option to specify the correct maximum length.

3) If the object ID is longer than what is allowed for a filename length, the object ID is transformed into a path by splitting the ID into filename max units until the ID is exhausted. Otherwise, two 1-character hash codes are developed from the object ID itself and each is used as a directory prefix to the object ID. The directory names are actually two character hexadecimal digits representing each hash character.

4) If a prefix is specified in the parms passed to the plugin, it is used to prefix the result from the previous step. This allows you to place all object ID’s in a separate directory in the cache.

Using the above steps, an object ID of say “foobar” would be placed in

/osslocalroot/prefix/68/7F/foobar

Where osslocalroot comes from the oss.localroot directive and prefix comes from parms passed to the plugin. Either or both may be null.

5.1.2 Cache Integrity Options

The following matrix shows how combinations of integrity options interact.

	cache	nocache
net	Verifies checksums or checksum equivalents when sent with data over the network and records them with the cached file. The checksum is verified when the data is read from the cached file.	Verifies checksums or checksum equivalents when sent with data over the network. No checksum integrity checks are performed on the cached file.
nonet	Does not perform a transmission integrity check but does generate unverified checksums on the received data and records them with the cached file. The checksum is verified when the data is read from the cached file.	Does not perform any integrity checking on received or read data. This combination is equivalent to specifying off.

Network data integrity is very much dependent on the data provider supporting checksums on transmitted data. All XRootD data servers, release 5 and above, support transmission integrity using crc32c checksums. If a server does not support checksum integrity but does support TLS, the data is transmitted using TLS to verify integrity and crc32c checksums are then locally generated unless notls has been specified. In either case, these are called verified checksums.

When a data server supports neither checksum integrity nor TLS, the data cannot be verified. However, if the cache option is selected checksums are still generated locally and recorded with the cached file so that data at rest integrity can be performed. While these checksums are unverified they still detect media data alteration when the data is read back; even if the data was transmitted with data errors. The same is true if the combination cache nonet is specified.

Since a file can come from multiple sources, some of which providing unverified data; a cached file can be composed of verified and unverified pages. The uvkeep option may be used to limit exposure to corrupt data. This option specifies the longest time a cached file containing unverified pages may be used to provide data; after which the cached file is deleted and the data is fetched again, if needed. This option is useful when dealing with a large number of unverifiable data sources.

When only a handful of data sources provide unverified data, the uvkeep 0 option can be used to ensure that a cached file never contains unverified data. Any data received without a checksum or checksum equivalent (i.e. TLS) is never written to media and is only served from memory. While this does increase overhead as data may need to be fetched multiple times, the overhead should be small if most of the data sources support data transmission integrity.

If the major problem is media corruption and virtually all network transmissions are error-free, the combination of cache nonet may make sense.

Finally, the cache integrity relies on the data source to ensure that correct data is actually being sent over the network. That means the data source should also have verified that the data has not been corrupted not only upon creation but also due to media failures when it is read back. The data integrity option provided by the caching proxy can only detect, at best, transmission and cache media errors.

6 Memory Caching Proxy

The memory caching proxy is almost identical to the standard proxy except that memory is used to hold recently read data. The memory cache is used to satisfy future reads if the requested data is present in the cache. This avoids re-fetching data from the originating server. A memory cache proxy may make LAN transfers more efficient, depending on the data access pattern.

You configure a memory caching proxy the same way you configure a standard proxy server with the addition of a pss.cache directive, described in the next section. Be aware that disk caching proxies and memory caching proxies are mutually exclusive (i.e. you cannot configure both in the same server).

6.1 Configuration

Proxy memory caching specific directives and options are described below.

pss.cache [cacheopts]

cacheopts: [debug dbg] [logstats] [max2cache mc] [minpages mp]

[pagesize ps][preread [prpgs [minrd]]

[perf prf [rcalc]]] [r/w]

[sfiles {off | on | .sfx}] [size sz] [cachepots]

Where:

cache

Configures memory caching of data. By default, memory caching is turned off. Specifying the cache directive with no cachepots is equivalent to specifying:

debug 0 minpages 256 pagesize 32k preread 0 sfiles off size 100m

If the cachelib directive is specified, the cache options have no effect. Otherwise, the following options may be specified:

debug sets the debugging level: 0 is off, 1 is low, 2 is medium, and 3 is high. The default is 0.

logstats prints statistics about cache usage for every file that is fully closed.

max2cache the largest read to cache in memory. The value is set to the pagesize if it is smaller than pagesize or not specified (i.e. default). The limit does not apply to pre-read operations. The mc value may be suffixed by k, m, or g to scale mc by 2¹⁰, 2²⁰, or 2³⁰, respectively.

minpages specifies the minimum number of pages that the cache should hold. If it is unspecified or is less than one it is set to 256.

pagesize the size of each page in the cache. It is adjusted to be a multiple of 4k. This also establishes the minimum read size from a data server. The ps value may be suffixed by k, m, or g to scale ps by 2¹⁰, 2²⁰, or 2³⁰, respectively. The default is 32K.

preread enables pre-read operations. By default no data is pre-read. Specify the number of additional pages to be read past the last read page. You also follow prpgs by the read size that triggers a pre-read. Reads less than minrd cause a pre-read to occur. The minrd value may be suffixed by k, m, or g to scale minrd by 2¹⁰, 2²⁰, or 2³⁰, respectively. The default is the pagesize value plus one. Specifying preread with no options is equivalent to specifying:

preread 1 perf 90

preread perf specifies the minimum required performance from automatic pre-reads and is part of the preread options. When this performance cannot be obtained for a file, pages are no longer automatically pre-read for the file. Performance is measured as the number pre-read pages used divided by the number actually pre-read times 100. Specify a value between 0 and 100 for prf. The default is 90 (i.e., 90% of all pre-reads must be useful). The prf may be followed by rcalc, the number of pre-read bytes that trigger a new performance calculation. The rcalc value may be suffixed by k, m, or g to scale num by 2¹⁰, 2²⁰, or 2³⁰, respectively. The default is 52428800 (i.e. 50m).

r/w caches files opened in read/write mode. By default, only files opened read/only are cached.

sfiles enables or disables optimization for structured file (e.g., root files). Cache usage is optimized for the typical access patterns associated with structured files. Specify off to turn off this feature (the default), on to always use this feature, or a dot followed by a suffix (i.e. .sfx). Only files whose names end with this suffix, including the dot, will have this optimization applied.

size specifies the size of the cache in bytes. See the notes on how this value is adjusted irrespective of the specification The sz value may be suffixed by k, m, or g to scale sz by 2¹⁰, 2²⁰, or 2³⁰, respectively. The default is 100m.

Notes

1) The size value may be adjusted upwards or downwards depending on the pagesize and minpages values. The following adjustments are made:

a. The size value is adjusted downwards to be a multiple of pagesize.

b. The final size value is either the adjusted size or minpages multiplied by the pagsize,,whichever is greater.

7 POSIX Server-less Caching

The subsystems that support caching proxy servers may also be used to setup a server-less proxy cache. This is essentially a client-side application cache. It caches data either in memory or on disk (depending on the configuration) by intercepting file system calls. Any data that is read is locally cached. Currently, this feature is available only via the POSIX preload library (i.e. libXrdPosixPreload.so) or by directly linking with libXrdPosix.so and using the XRootD POSIX interface API’s defined in one of the following public include files:

· XrdPosix.hh,

· XrdPosixExtern.hh, or

· XrdPosixXrootd.hh

Using the XRootD POSIX API’s is straightforward as they do not differ from the standard POSIX API’s. The API’s defined in XrdPosix.hh and XrdPosixExtern.hh accept either simple paths or XRootD URL’s. A simple file path vectors the request to the local file system while an XRootD URL (i.e. one that starts with xroot:// or root://) vectors the request to the specified server. You can define virtual mount points that convert specified simple paths to actual URL’s. The interface defined in XrdPosixXrootd.hh only accepts XRootD URL’s.

The preload library mechanism offers a very simple way to run POSIX-compliant programs without needing to change them or even recompile them. You can invoke your program with a simple shell script shown below.

LD_PRELOAD=/usr/lib64/libXrdPosixPreload.so

export LD_PRELOAD

Note that you may need to change the LD_PRELOAD value if the preload library is not installed in a standard place. So, assume the script is called runit then executing

runit ls /tmp

produces a directory listing of the local /tmp directory while

runit ls xroot://someserver//tmp

produces a directory listing of someserver’s /tmp directory, if allowed.

Regardless of whether you use the direct API or the preload library, you need to tell the XRootD POSIX subsystem that you wish to cache data (i.e. configure caching).

This is done by setting the XRDPOSIX_CONFIG environmental variable to the path to your configuration file. For instance, we can augment the previous runit script to also specify the location of the configuration file that can be used to enable caching.

LD_PRELOAD= /usr/lib64/libXrdPosixPreload.so

XRDPOSIX_CONFIG=/home/abh/psx.cf

export LD_PRELOAD XRDPOSIX_CONFIG

So, /home/abh/psx.cf contains the needed directives to setup local caching.

7.1 Server-less Disk Caching

To enable disk caching, you need to specify at least the following directives in your configuration file.

posix.cachelib /usr/lib64/libXrdPfc.so

oss.localroot cachepath

pfc.diskusage fracLow fracHigh

pfc.ram bytes[m|g]

The cachelib directive tells the system to use local disk caching using the libXrdPfc.so plug-in. You may need to change the path is this plug-in is installed in a non-standard directory.

The localroot directive specifies the disk location of the cache. The cachepath should be some existing directory path where cached file blocks can be permanently placed.

The diskusage directive specifies how much disk to use and at what point unused cached blocks should be deleted. Specify for fracLow the space utilization percentage that triggers purging. Specify for fracHigh the percentage of space utilization that must be reached for purging to stop. The default is “0.9 0.95” (i.e. 90% 95%) but a more reasonable combination might be “0.45 0.50” to prevent total use of your local disk. The diskusage boundaries, can be specified also in g or t bytes units for giga- and tera-bytes, respectively. Since controlling the amount of disk space that the file cache can use is tricky, you may wish to use a separate disk partition for the cache.

Finally, the ram directive tells the disk cache the maximum amount of memory that it may use. The larger the amount the better will be the performance.

All of the “pfc.” prefixed directives are available for your use (see the Disk Caching Proxy Configuration section). A limited number of ”pss.” directives are also available but these directives must be prefixed with “posix.” in order to be recognized (see the section on esoteric directives). Since the disk caching subsystem uses the standard storage system plug-in to handle the disk cache, you can also use most of the “oss.” prefixed directives documented in the “Open File System & Open Storage System Configuration Reference”. However, it’s best to ignore the “pss.” and “oss.” directives unless you have very special needs.

7.2 Server-less Memory Caching

Using memory caching is far simpler than using disk caching because there is less to configure. Your configuration file needs to only have one directive, as shown below.

posix.cache [cachepots]

The cacheopts are identical to the ones described in the Memory Caching Proxy Configuration section using the pss.cache directive.

7.3 Esoteric Directives

The following table lists available directives that enable highly specialized features. For details on these options see the Proxy Configuration section.

POSIX Directive	Documented	Notes
posix.ccmlib	pss.ccmlib	Allows running a cache context manager.
posix.ciosync	pss.ciosync	Controls cache synchronization on close.
posix.inetmode	pss.inetmode	Controls TCP/IP stack usage.
posix.namelib	pss.namelib	Allows name mapping and cache squashing.
posix.setop	pss.setopt	Controls.underlying XRootD client.
posix.trace	pss.trace	Enables tracing.

7.4 Defining Virtual Mount Points

When using the POSIX preload library or the XrdPosixExtern.hh interface, you can define virtual mount points. A virtual mount point equates a path to an XRootD URL. Thus, any use of that path is internally converted to a specified URL using a template before being handed to the XRootD POSIX subsystem.

You specify virtual mounts by setting the XROOTD_VMP environmental variable to one or more templates that specify how to convert a path to a URL. The basic format is shown below.

server[:port]:path[=[newpath]] [ . . . ]

Where:

server is the DNS name or IP address of the target server that would have been specified in a URL.

port is the optional port number. If not specified, 1094 is assumed

path is the path prefix that is associated with server. It must start with a slash. Any specified path whose prefix matches path causes the specified path to be prefixed with “xroot://server[:port]/” (i.e. an XRootD URL) and then vectored to the server. This subject to what has been specified for the remainder of the template.

= when a simple equals sign follows path, The path prefix (i.e. path) is removed from the specified path before prefixing it with an XRootD URL.

newpath

when specified after the equals sign, newpath replaces path before prefixing it with an XRootD URL.

You may specify any number of virtual mount point templates in the environmental variable. Each one must be separated by a single space.

7.4.1 Examples

XROOTD_VMP=xnode:/xrootd/

/xrootd/home/abh is converted to xroot://xnode:1094//xrootd/home/abh

XROOTD_VMP=xnode:2094:/xrootd/

/xrootd/home/abh is converted to xroot://xnode:2094//xrootd/home/abh

XROOTD_VMP=xnode:/xrootd/=/

/xrootd/home/abh is converted to xroot://xnode:1094//home/abh

XROOTD_VMP=xnode:2094:/xrootd/=/atlas/

/xrootd/home/abh is converted to xroot://xnode:1094//atlas/home/abh

9 Document Change History

16 Jun 2014

· Split description of proxy services from the ofs/oss manual. Hence, this is a new manual.

· Document the caching proxy service.

29 Jul 2014

· Document forwarding proxies.

· Document the pss.origin directive that configures a forwarding proxy.

5 Aug 2014

· Document the pss.permit directive that restricts outgoing connections of a forwarding proxy.

2 Dec 2014

· Document the disk caching proxy configuration directives (i.e. the ones that start with pfc).

14 Feb 2015

· Change the “pfc.hdfs” prefixed directives to “pfc.file” prefixed directives.

· Remove the pfc.chachedir directive.

11 Oct 2015

· Clarify the meaning of the high and low watermarks in the pfc.diskusage directive .

11 Jan 2017

· Document version 2 of the disk caching proxy which includes new and changed directives.

21 Mar 2017

· Correct pfc.ram directive (was incorrectly pss.ram).

· Document the -lfncache and -lfn2pfn pss.namelib directive options.

21 May 2017

· Describe automatic proxy selection via a client plug-in.

· Describe how disk caching proxy overloads may be handled.

26 June 2017

· Describe how to cache S3-type objects.

28 June 2017

· Describe the pss.ciosync directive.

· Document the pss.config directive’s stream option.

· Document server-less caching

· Document XRootD POSIX interface.

18 September 2018

· Describe the enhanced pfc.diskusage directive.

16 May 2019

· Describe the pfc.writequeue directive.

· Describe the pss.ccmlib directive.

· Describe the pss.dca directive.

· Add section on Direct Cache Access.

21 January 2020

· Describe URL specification as the origin server using the pss.origin directive.

27 March 2020

· Document the pss.persona directive.

· Restructure the configuration section for readability.

18 May 2020

· Document that forwarding specification of the pss.origin directive may accept additional protocols that are to be considered acceptable.

2 June 2020

· Document the pfc.acchistorysize, pfc.flush, and the -lfncachesrc option of the pss.namelib directive.

· Expand security section and add privacy and integrity section.

· Simplify cluster configuration using the cache export attribute.

24 August 2020

· Correct documentation for the pss.dca directive.

31 August 2020

· Correct documentation for the pfc.cschk directive.

9 September 2020

· Correct documentation for the pfc.cschk directive and add a section describing data integrity options.

6 November 2020

· Describe the [no]tls option on the pfc.cschk.

21 June 2021

· Describe the pss.reproxy.

· Change the pfc default block size to 256K.

· Add description of the interaction between the pfc.blocksize and the xrootd.async directives.

7 June 2022

· Change the pfc default block size to 128K.

1.2 Forwarding Mode Proxies

2 Security Considerations

4.3 Configuration

4.3.1.1 Specifying the origin as a URL

5 Disk Caching Proxy

5.1 Configuration

5.1.2 Cache Integrity Options

5.3 Handling Cache Overloads

6 Memory Caching Proxy

6.1 Configuration

7.3 Esoteric Directives

7.4 Defining Virtual Mount Points