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. Collectively, this is 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.

By default, all special features are disabled. A progression of directives enables certain features. Features are enabled as described in the following table.

Directive	Purpose
ofs.osslib	Load the shared library implementing a special pss (storage system) component.

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.

1.2.3 Security Considerations

Proxy servers never forward client credentials to the endpoint that they contact. Instead, they use their own credentials to establish authenticity. This works well for incoming connections as the proxy can validate the client before performing any actions on behalf of the client. However, outgoing connections pose a problem; especially for forwarding proxies which are meant to be used for outgoing connections. Validating a client using internal site rules does not necessarily capture any specific access restrictions in effect outside the site. For instance, a site may authorize an outgoing client but that client may not be actually authorized by the final endpoint. If the proxy itself is authorized by the final endpoint then, in effect, the client gets access to resources that would otherwise be prohibited if the client made direct contact with the outside endpoint. Consequently, great care must be given when providing resources using a forwarding proxy.

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.

1.4 Minimal Sample Configuration Files

# Specify that we are a direct mode proxy fronting the host

# data.stanford.edu (server or redirector)

pss.origin data.stanford.edu

# The export allows access to any path via proxy as the

# origin host will enforce its own exports

all.export /

# We need to load the proxy plugin for this to actually work

ofs.osslib libXrdPss.so

Minimal Direct Mode Proxy Configuration

# Specify that we are a forwarding proxy

pss.origin =

# The export allows xroot and root type URL’s destinations

all.export /xroot:/

all.export /root:/

# We need to load the proxy plugin for this to actually work

ofs.osslib libXrdPss.so

Minimal Forwarding Mode Proxy Configuration

# Specify that we are a forwarding proxy for URL type of paths

# and a direct mode proxy for any other type of path

pss.origin = data.stanford.edu

# The export allows any path to be specified. However, paths

# starting with “/xroot:/” or “/root:/” are treated as URLs

# and are forwarded. Anything else is accessed via

# data.stanford.edu

all.export /

# We need to load the proxy plugin for this to actually work

ofs.osslib libXrdPss.so

Minimal Combination Mode Proxy Configuration

3.2 Proxy Cluster

A proxy cluster consists of two or more simple proxy servers arranged as an xrootd cluster. While one could use a load-balancing DNS to bridge together multiple proxy servers, using the xrootd cluster management infrastructure to do this provides not only true load balancing but also fail-over. That is, should a proxy server become unresponsive an xrootd client will automatically switch to another working proxy server. This also allows you to take down proxy servers for maintenance without worrying if they are currently in use, as long as one proxy server remains.

The mechanics of setting up a proxy cluster is largely the same as setting up a regular xrootd cluster. However, different parameters are specified for the all.role and all.manager directives. Additionally, unlike a simple proxy server you must also run a cmsd daemon everywhere you run an xrootd proxy server and you must configure a proxy redirector. The initial point of contact for external clients is the proxy redirector. This is shown in the adjacent diagram.

Setting up a proxy cluster is a minor variation of setting up a standard xrootd cluster. For proxy servers you need to identify the proxy manager and tell each proxy server that it is part of a proxy cluster. This allows the proxy servers to cluster around the proxy manager and the proxy manager then knows how to redirect clients to a working proxy server. A sample configuration file follows. You should compare it to one for a simple xrootd data cluster.

All red directives are required. The all.export and pss.origin directives are the same as for simple proxy servers. These tell the xrootd’s what to expor, and where the xrootd data cluster resides, respectively.

The all.manager directive tells each proxy server who the proxy redirector is and what port the cmsd is using (here we arbitrarily chose 1213).

In the if-fi construct roles are assigned to each server. Since z.domain.edu is the redirector, its role is proxy manager. All other nodes in the proxy cluster have a role of proxy server. This also causes the ofs layer to automatically load libXrdPss.so, by default, as the ckslib and the osslib; making the ofs.ckslib and ofs.osslib directives unnecessary. If the defaults are inappropriate, specify these directives to override the default.

Since we only want true proxy servers to be allowed to connect to the cmsd running on the proxy redirector node, the second if-fi construct specifies that only those nodes are allowed to connect. This is especially important because the proxy redirector is publically accessible and cmsd connections should be restricted to a known set of hosts. The particular example uses a generic specification. You may wish to make “allow” more specific. For instance, if proxy servers have a DNS name of “proxynn.domain.edu”, where nn is a sequence number, then it would be better to specify “xrd.allow host proxy*.domain.edu”.

The green directives, xrd.allow and sec.protocol, are used to limit who can use the proxy servers. 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. You should especially note that these directives apply only to the xrootd and not the cmsd. This is why they have been qualified by an if-fi clause that applies them only to the xrootd proxy servers.

The default load distribution is to round-robin requests across all available proxy servers. You can use the cms.sched directive for the proxy redirector along with the cms.perf directive for proxy servers to do load balancing on actual load. These directives are described in the cmsd reference manual.

When configuring a proxy cluster, all proxy servers must either be in direct mode, forwarding mode, or combination mode. Any attempt to mix modes will lead to unpredictable result.

If you are using disk-caching proxies, a different type of cluster configuration must be used. Refer to the description of disk caching proxy clusters for details.

3.3 Proxy Server Directives

The proxy server plug-in is based on xrootdFS which, is based on the xrootd POSIX library which, in turn, uses the xrootd client library. Each of these components has specific options that can be specified in the same configuration file used for the proxy setup. Specifically, the following options are processed by the proxy

· all.export

· oss.defaults

Proxy specific directives and options are described below.

pss.ciosync ssec msec

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

pss.dca [recheck {tm | off}]

pss.inetmode {v4 | v6}

pss.localroot path

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

pss.origin dest | = [dest] required directive

pss.permit [/] [*] hspec

pss.setopt ConnectTimeout seconds

pss.setopt DataServerConn_ttl seconds

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

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

dest: host[:port | port]

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.

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:

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

inetmode

Specifies the 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.

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.

-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.

origin specifies the location of the redirector or server for which the server being configured is to proxy. This directive is required.

host The DNS name or IP address of the redirector or server being proxied.

port The TCP port number or service name of the redirector or server associated with host. The port may be specified with an adjacent colon or space separation. The default is 1094.

= Configures the proxy in forwarding mode where the client supplies the actual endpoint. If a host and optional port follows the equals sign then this endpoint is used when the client has not specified a root or xroot URL as an endpoint. Otherwise, the client receives a “not supported” error.

permit

specifies an allowed destination that a client can use via a forwarding mode proxy. If no permit directives are specified, no restrictions apply. Permits, by default, apply to path and objectid URLs.

/ The permit applies only to path-type URLS.

* The permit applies only to objectid-type URLS.

hspec The DNS host name allowed as a target connection. Substitute for name a host name or IP address. A host name may contain a single asterisk anywhere in the name. This lets you allow a range of hosts should the names follow a regular pattern or if you wish to allow connections to all hosts in a particular domain. IP addresses may be specified in IPV4 format (i.e. “a.b.c.d”) or in IPV6 format (i.e. “[x:x:x:x:x:x]”).

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.

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.

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 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.

2) 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.

3) For more information on the pss.dca directive see the section on Direct Cache Access

4 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 libXrdFileCache.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.

4.1 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.

pss.cachelib path [libopts] required directive

pss.ccmlib path [libopts]

pfc.blocksize bytes[k|m]

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.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

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 1m (i.e. one megabyte).

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.

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.

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.

Examples

a) Enable proxy file prefetching:

pps.cachelib libXrdFileCache.so

pfc.ram 100g

oss.localroot /data/xrd-file-cache

pfc.blocksize 512k

pfc.prefetch 8

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

pss.cachelib libXrdFileCache.so

pfc.ram 100g

oss.localroot /data/xrd-file-cache

pfc.hdfsmode hdfsbsize 64m

6 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.

6.1 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/libXrdFileCache.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 libXrdFileCache.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.

6.2 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.

6.3 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.

6.4 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 “xrootd://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.

6.4.1 Examples

XROOTD_VMP=xnode:/xrootd/

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

XROOTD_VMP=xnode:2094:/xrootd/

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

XROOTD_VMP=xnode:/xrootd/=/

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

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

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

1 Introduction

1.1 Direct Mode Proxies

1.2 Forwarding Mode Proxies

1.2.1 Forwarding File Paths

1.2.2 Forwarding Object IDs

1.2.3 Security Considerations

1.3 Combination Mode Proxies

1.4 Minimal Sample Configuration Files

2 Automating Proxy Service Selection

3 Standard Proxy Service

3.1 Simple Proxy Server

3.2 Proxy Cluster

3.3 Proxy Server Directives

4 Disk Caching Proxy

4.1 Configuration

4.1.1 Disk Caching S3-type Objects

4.2 Disk Caching Proxy Clusters

4.3 Handling Cache Overloads

4.4 Direct Cache Access

5 Memory Caching Proxy

5.1 Configuration

6 POSIX Server-less Caching

6.1 Server-less Disk Caching

6.2 Server-less Memory Caching

6.3 Esoteric Directives

6.4 Defining Virtual Mount Points

6.4.1 Examples

7 The netchk Utility

8 Document Change History