dc_test (1) - Linux Manuals

dc_test: Distributed session cache testing and benchmarking tool

NAME

dc_test - Distributed session cache testing and benchmarking tool

SYNOPSIS

dc_test -connect <addr> [options]

DESCRIPTION

dc_test runs a configurable number of session caching operations against a given network address using the distcache(8) protocol. This is typically to verify an installation, to help diagnose problems, and to measure latency and/or throughput. The given address could be for an instance of dc_client or directly to a running instance of dc_server. Note that whenever the service is running over TCP/IPv4, use of the -persistent flag is advised because many systems will fairly quickly exhaust the limit of available file-descriptors and stop allowing new sockets until the closed connections finish with the TIME_WAIT state.

dc_test runs one cache operation at a time, so it provides an accurate measurement of latency in the distcache architecture being tested. However if one wants to test the potential throughput of the architecture when multiple cache operations occur in parallel (and distcache tools are asynchronous to scale well in such circumstances), it makes sense to launch multiple instances of dc_test in parallel.

dc_test attempts to verify correctness of any optimisation caching tricks in distcache tools by creating an array of ``sessions'' in advance and then sending sending ADD, GET, and REMOVE requests on these sessions chosen completely at random. As such, many of these requests should fail, and dc_test will itself immediately halt with an error if any operation that should succeed fails or any operation that should fail succeeds.

OPTIONS

-connect address
This flag specifies the address dc_test should connect and send cache operation requests to.
-progress num
Normally dc_test will remain silent while testing is in progress and will only produce output when complete (whether it is because an error was detected or because the required number of operations have completd). With this flag, a line of output is produced each time another num operations have been completed. This can be useful for giving an intuitive idea of the latency in the architecture and to verify that neither dc_test nor the tools it connects to have deadlocked.

The default value for this flag is 0, which leaves dc_test silent until it has completed.

-sessions num
This flag controls the number of sample sessions created by dc_test before testing begins. The default value is 10.
-datamin num
-datamax num
The sessions created by dc_test during initialisation have random session IDs between 10 and 63 bytes in length, and random session data of varying length also. The minimum length of the session data is set by the -datamin switch, and the maximum length is set by the -datamax switch. The default value for -datamin is 50, and the default for -datamax is 2100.
-timeout secs
-timevar secs
The sessions created by dc_test during initialisation have random timeouts associated with them, so as to better probe for potential expiry-ordering bugs in a cache server. Each session addition to a cache server must specify an expiry period after which the cache will be automatically removed if it hasn't already been removed explicitly or implicitly because of the FIFO behaviour of a full cache. The -timeout switch specifies the average timeout for creating sessions with, and the -timevar switch specifies the range above or below the average that session timeouts should be (randomly) chosen from.

The default value for -timeout is 60 seconds, and the default value for -timevar is 5 seconds.

If dc_server is not running as a daemon, then it can also be useful to watch it's behaviour after testing has finished (or in the middle of subsequent testing) because it should automatically scroll out test sessions as their expiry period runs out. See also the BUGS section.

-ops num
dc_test does not (yet) support running indefinitely (see the BUGS section for the primary reason), so it will run until any error has been observed or until it has successfully performed a configurable number of operations. The -ops switch can be used to set this number of operations.

If -ops is not supplied on the command-line, it's default value will be calculated from the number of sessions created (which is itself controlled by -sessions). This default is 10*(sessions^2). As the default number of sessions is 10, the default number of operations is thus 1000.

-persistent
The default behaviour of dc_test is to open and close a connection to the target address (controlled by -connect) for every operation. This is the logical choice when connecting to an instance of dc_client as it is the mode of operation employed by most applications that integrate with distcache. However, it can be desirable to measure the behaviour of the architecture when all operations are taking place over a single established ``persistent'' connection. This is more certainly the case when connecting directly to an instance of dc_server over TCP/IPv4, as dc_server forks a child process for each connection and TCP/IPv4 does not react well to thousands of connections being opened and closed in a short space of time.

Using the -persistent switch, a single connection will be opened to the target address when dc_test initialises, and all cache operations will use this connection.

-h, -help, -?
Any of these flags will cause dc_test to display a brief usage summary to the console and exit cleanly. Any other flags are ignored.

BUGS

dc_test will monitor each of the sessions it creates during intialisation to determine whether it thinks it should already be present in the destination session cache or not. Using this state, it determines whether it expects a corresponding ADD, GET, or DELETE operation on the session to succeed or fail. GET and DELETE should fail if the session is not supposed to be in the cache (if it has never been added or was last seen in a DELETE operation). Likewise, ADD should fail if the session is supposed to be in the cache. None of this logic, however, pays any attention to the expiry period in each of these sessions and so if a long enough test is run (high value of ops or a slow architecture) and/or short session timeouts are used, it becomes possible for this state to become incorrect when the cache itself expires sessions that dc_test thinks should still be in the cache. This bug should be hard to trigger even with long tests or short timeouts, because a session needs to be added to the cache and then not DELETE'd during a long enough period that the server gets a chance to expire it.

AUTHOR

This toolkit was designed and implemented by Geoff Thorpe for Cryptographic Appliances Incorporated. Since the project was released into open source, it has a home page and a project environment where development, mailing lists, and releases are organised. For problems with the software or this man page please check for new releases at the project web-site below, mail the users mailing list described there, or contact the author at geoff [at] geoffthorpe.net.

Home Page: http://www.distcache.org

SEE ALSO

dc_client(1)
Distributed caching client proxy.
dc_server(1)
Distributed cache server.
dc_snoop(1)
Distcache protocol analyser and debugging tool.
distcache(8)
Overview of the distcache architecture.
http://www.distcache.org/
Distcache home page.