<html> <head> <link rel=stylesheet href="style.css" type="text/css"> <title>collectl - Lustre</title> </head> <body> <center><h1>Lustre</h1></center> <p> <h3>Overview</h3> The first thing to understand about lustre reporting is in most cases, where one has configured the server(s) and just wants to monitor them, all one need do is specify -sl or -sL and collectl will do the right thing. It will automatically detect the type of service(s) currently running and will either record or display the appropriate data. If you select -sl and the system doesn't have lustre installed, it will warn you and then disable that switch. <p> <h3>Controlling Which Data is Displayed</h3> Lustre records a wealth of performance data, far more than makes sense to display all the time, and so by default collectl displays minimal information such as bytes/operations read and written. At the client detail level lustre can differentiate this data at the filesystem and even the OST level! In order to accommodate the broadest flexibility one is allowed to control the way data is collected/displayed via several complementary switches. <ul> <li>-s: As is normally the case, one can specify '-sl' for summary level data, '-sL' for detail data or combine them to get both. However, since the client detail data can actually be presented at the individual filesystem or OST level, there is an option to show the OST level details (filesystem details are the default) <i>see --lustopts O</i>. <li>--lustopts: This switch is used to provide further detail about the types of data that is to be collected/displayed. There are 5 such values that collectl cares about: <ul> <li>B - rpc buffer level data.</li> <li>D - disk block statistics, which applies to both MDS and OSS servers. One should also note this is specific to HP SFS and this data is not available in the open source version.</li> <li>M - client metadata (note that this was the default prior to collectl V1.6.2).</li> <li>O - for client details only, show results by OST <li>R - read_ahead statistics. Unlike the other options, which generate a lot of data, <i>--lustopts R</i> may be used with brief mode.</li> <p> As it turns out, nothing is quite as simple as it seems and while the following case is not typical, it needs to be addressed for completeness. Since collectl allows one to collect one set of data and to later display a different set, consider what happens in one were to collect multiple types of lustre data for a client using <i>--lustopts MR</i>, but then just play back the basic client data which is collected without specifying <i>--lustopts</i>. By default, playback mode defaults to the settings data was collected with and to change the display one needs to explicitly change those settings. To meet this need, use <i>--lustsvc</i>, which is described in more detail later. </ul> </ul> <p> In the spirit of letting the user display whatever they want to, collectl will allow one to select multiple values for <i>--lustopts</i> and it will try to display the results appropriately. Perhaps the easiest thing to do is just experiment and in most cases you'll get what you're looking for. There are a few combinations of <i>-s</i> and <i>--lustopts</i> that do not make sense and if you choose one, you will be told. <p> <h3> What About Playback?</h3> As is always the case with playback, unless otherwise told to do something else, collectl will playback its recorded data based on the parameters selected for collection. In other words, if you specify <i>--lustopts OBR</i> in record mode, collectl will record both RPC buffer and read_ahead stats. When you play the data back, it will then display both as well. However, you also have the option of specifying <i>--lustopts</i>, typically thought of as a collection-only switch, and it will force the output to what you'd like it to be. If you select a statistics type that hasn't been recorded, that information will be displayed, but as zeros. <p> <h3>Recognizing Service Configuration Changes</h3> In some cases lustre services may change after collectl starts. In fact, it may not even be running and if so you'll get a message telling you it is not and that collectl cannot determine the system type since it could be a client, MDS, OSS or some combination. This includes services starting and stopping as well as the configurations of those services themselves changing. For example one might occasionally mount/umount different lustre filesystems on a client. Not to worry. Collectl periodically checks for configuration changes and automatically adjusts the data it collects as well as anything it may be currently displaying. However this can also lead to the output format changing. If you know that the system type could change and you simply want to force the type of output to be consistent, use <i>--lustsvc</i> as described in the next section. <p> <h3>Changing the Default Recording/Display Behavior</h3> There are some times when you want specific control over what data is recorded or displayed rather than the default behavior OR collectl starts before lustre does and it can't determine the type of system it is. This is typically the case when a system is playing multiple roles by providing more than one service. For example, if a system has been configured as both an OSS and a client, every time you run collectl you will collect or display data about both and sometimes this is NOT what you want. There may be other times where you have developed some reports or graphs that expect data in a standard format and you've collected a subset (or superset) of data. <p> To override this behavior of the lustre portion of the data (remember you can control the displaying of individual subsystems with -s), use <i>--lustsvc</i> to specify the type of service(s) you're interested in and collectl will only pay attention to those, both for recording to a file as well as display. Naturally when displaying data for services you never collectled data on, those services will print as zeros. <p> If all this sounds confusing, just experiment with various combinations of <i>-s, --lustopts</i> and <i>--lustsvcs</i> and observe the behavior. <table width=100%><tr><td align=right><i>updated Mar 26, 2010</i></td></tr></colgroup></table> </body> </html>
