Tutorial - Lustre


This tutorial is intended to help you get started monitoring a system on which the lustre filesystem has been installed. This is not intended to be a tutorial on how to use such a filesystem. In terms of the basics, there's really not much to say other then tell collectl to monitor lustre by specifying an l with -s along with any other subsystems you may want to monitor.

As you should be aware of by now, collectl will try to display everything you've selected in brief format if it can, writing all your data out on a single line for each sampling interval. This can get quite wide depending on how many subsystems you choose to monitor and while you can cerainly specify collectl -s+l to add lustre to the default subsystems, the output width is too cumbersome for this tutorial. Therfore I'll use some other subsystems and switches in the examples to mix things up, showing that there are a lot of possible combinations. In this first example, we see what happens when you run collectl on a lustre client and request cpu and memory data along with lustre.

$ collectl -scml
#<--------CPU--------><-----------Memory----------><-------Lustre Client------>
#cpu sys inter  ctxsw free buff cach inac slab  map  Reads KBRead Writes KBWrite
   0   0   101     26   3G   6M  29M  12M  43M  65M      0      0      0       0

Collectl is actually very intelligent about dealing with lustre because if you were to run the identical command on an OST, it would recognize that too and change what it shows accordingly as you can see below.

$ collectl -scdl
#<--------CPU--------><-----------Disks-----------><--------Lustre OST------->
#cpu sys inter  ctxsw KBRead  Reads  KBWrit Writes KBRead  Reads KBWrit Writes
   0   0   100     28      0      0       0      0      0      0      0      0
In fact, if you're also running a client on an OST it will show both! I've also included time stamps to make the output a little more intersting.
$ collectl -scl -oT
#         <--------CPU--------><--------Lustre OST-------><-------Lustre Client------>
#Time     cpu sys inter  ctxsw KBRead  Reads KBWrit Writes  Reads KBRead Writes KBWrite
14:35:32    0   0   103     24      0      0      0      0      0      0      0       0
14:35:33    0   0   123     53      0      0      0      0      0      0      0       0
As expected, you can run collectl on a system that has any combination of MDS, OST and client services and it will show you what you want to see. For more detail on some of the more advanced concepts beyond the scope of this tutorial you can read more about how collectl deals with Lustre here.

And finally, don't forget any of this data can be written to a file for continuous logging, played back later and even converted to a format suitable for plotting. In fact collectl is configured to monitor lustre by default when run as a daemon so all you need to do to begin collecting the basic data shown above is service collectl start.

Beyond the Basics

For many users of collectl, you are now sufficiently equipped to tackle most lustre monitoring tasks, but for those more exotic situations there's so much more you can do.


Let's start off by looking more closely at client data. As with all other collectl data, one can switch between summary and detail data by simply entering an upper case L instead of a lower case one as I've done below. The actual content of what is displayed will depend on whether you are on a client, OSS (there is currently no detail data for an MDS), but as you can see for clients, this data is broken down by filesystem and just to make the display a little more interesting I decided to show the timestamps in milli-seconds. Naturally if there is only one filesystem, the data with match that displayed in summary mode.

$ collectl -sL -oTm
#            Filsys   Reads ReadKB  Writes WriteKB
15:10:06.009 spfs1       0      0       0       0
15:10:06.009 spfs2       0      0       0       0
Just to take it one step futher, it turns out that lustre actually tracks client I/O by individual OSTs and sometimes that is more interesting, so there is a special option for lustre clients, --lustopts O.
$ collectl -sL --lustopts O -oTm
#            Filsys  Ost      Reads ReadKB  Writes WriteKB
15:19:17.007 spfs1  OST0000      0      0       0       0
15:19:17.007 spfs1  OST0001      0      0       0       0
15:19:17.007 spfs2  OST0000      0      0       0       0
15:19:17.007 spfs2  OST0001      0      0       0       0
So what else can we look at? Lots! Lustre tracks readahead data which you can show in brief format like this by using the R option noting this time we're also requesting date/time stamps be included. Here we see the cache hits/misses added to the brief display for the client.
$ collectl -sl --lustopts R -oD
#                  <-------------Lustre Client-------------->
#Date    Time       Reads KBRead Writes KBWrite   Hits Misses
20080319 15:20:12       0      0      0       0      0      0
20080319 15:20:13       0      0      0       0      0      0
or you can look at it in verbose format like this:
$ collectl -sl --lustopts R --verbose
# Reads ReadKB  Writes WriteKB  Pend  Hits Misses NotCon MisWin LckFal  Discrd ZFile ZerWin RA2Eof HitMax
      0      0       0       0     0     0      0      0      0      0      0      0      0      0      0
      0      0       0       0     0     0      0      0      0      0      0      0      0      0      0
Lustre also tracks client metadata, which you can request by specifying --lustopts M and BRW stats which are selected by --lustopts B. However both are so wide they only have a verbose form and the following shows both being displayed at the same time.
$ collectl -sl --lustopts BM
#Rds  RdK   1P   2P   4P   8P  16P  32P  64P 128P 256P Wrts WrtK   1P   2P   4P   8P  16P  32P  64P 128P 256P
   0    0    0    0    0    0    0    0    0    0    0    0    0    0    0    0    0    0    0    0    0    0
# Reads ReadKB  Writes WriteKB  Open Close GAttr SAttr  Seek Fsynk DrtHit DrtMis
      0      0       0       0     0     0     0     0     0     0      0      0
And if that's not enough, these even have a detailed mode of display and you can look at them by filesystem or OST. You can specify any combinations of B,M and R with --lustopts to show the associated data in both summary or detail formats, though only readahead appears in the brief display. All others will force verbose format.

Object Storage Server

If you haven't figured it out yet, it's also possible to display OST level information on an OSS by simply using the uppercase subsystem specification as you can see here, noting just to be different I've chosen a second form of the date/timestamp.

$ collectl -sL -od
#              Ost            Read Ops   Read KB      Write Ops   Write KB
03/19 15:32:14 spfs1-OST0000        0         0              0          0
03/19 15:32:14 spfs2-OST0000        0         0              0          0
03/19 15:32:15 spfs1-OST0000        0         0              0          0
03/19 15:32:15 spfs2-OST0000        0         0              0          0
You can also show BRW stats as both summary and detail data as well and since they look identical to the way they're displayed for a client, I won't bother repeating those forms here.

Metadata Server

As mentioned earlier, there is no detail data for an MDS nor are there any other types of data other than that which can be displayed in summary mode.


As you have seen, there is a wealth of data here and often you may not even know what you want to look for. When such is the case, just collect as much as you can and save it in a file. Then you can play it back later and display it in multiple formats.

I just want to close with an example of a problem with readaheads and Lustre 1.4 which demonstrates the power of a tool like collectl that can show multiple types of data at once because I know a lot of people may still not be convinced. Consider the following sample which was collected while doing random 32KB reads of a large file. See anything wrong? Do you know why the network bandwidth is so much higher than the lustre client read rate? How long would it have taken to even realize there is a problem?

$ collectl -snl -oT
#         <----------Network----------><-------------Lustre Client-------------->
#Time     netKBi pkt-in netKBo pkt-out  Reads KBRead Writes KBWrite   Hits Misses
08:14:18   41776  28310   1065   14786     50    200      0       0     30     20
08:14:19   38328  25987   1032   14078     62    248      0       0     35     19
08:14:20   44763  30337   1167   16114     58    232      0       0     30     20
08:14:21   43666  29596   1137   15632     46    184      0       0     30     16
08:14:22   33777  22905    891   12191     58    232      0       0     35     23
It turned out the old algorithm triggered a readahead after 2 consecutive pages were read and every 32KB read (which is 8 pages) resulted in 1MB being read over the network, loaded in to cache and was then discarded! If the value of max_read_ahead_mb is set to 0 for the associated filesystem on each client involved, you see 2 things - the network rate drops to track the client read rate and the misses goes to zero.
$ collectl -snl -oT
#         <----------Network----------><-------------Lustre Client-------------->
#Time     netKBi pkt-in netKBo pkt-out  Reads KBRead Writes KBWrite   Hits Misses
08:21:22       0      3      0       3     59    236      0       0      0      0
08:21:23     317    335     58     298     91    364      0       0      0      0
08:21:24     442    457     77     388    107    428      0       0      0      0
08:21:25     442    457     77     383     97    388      0       0      0      0
08:21:26     432    446     75     373     89    356      0       0      0      0
Caution - changing the value of the readable variable in /proc to 0 will eliminate all readahead for all readers meaning any applications that do sequential reads could have significant performance problems. In other words, this is not a recommendation to turn off readahead but rather to be aware of what it does and if you do choose to turn it off to be aware of the consequences

Note: This readahead algorithm has changed with V1.6 and lustre no longer triggers readahead on the third page read but rather on the third sequential read.
updated Mar 25, 2010