[lustre-discuss] Hidden QoS in Lustre ?

Tung-Han Hsieh thhsieh at twcp1.phys.ntu.edu.tw
Tue Oct 27 21:08:23 PDT 2020

Dear All,

Sorry that I am not sure whether this mail was successfully posted to
the lustre-discuss mailing list or not. So I resent it again. Please
ignore it if you already read it before.


Dear Andreas,

Thank you very much for your kindly suggestions. These days I got a chance
to follow your suggestions for the test. This email is to report the results
I have done so far. What I have done were:

1. Upgrade one client (with Infiniband) to Lustre 2.13.56_44_gf8a8d3f
   (obtained from github). The compiling information is:

   - Linux kernel 4.19.123.
   - Infiniband MLNX_OFED_SRC-4.6-
   - ./configure --prefix=/opt/lustre \
                 --with-o2ib=/path/of/mlnx-ofed-kernel-4.6 \
                 --disable-server --enable-mpitests=no
   - make
   - make install

2. We mounted the lustre file system (lustre MDT/OST servers: version
   2.12.4 with Infiniband with ZFS backend) by this command:

   - mount -t lustre -o flock mdt at o2ib:/chome /home

3. The script to simulate large data transfer is following:
   (the directory "/home/large/data" contains 758 files, each size 600MB)

   while [ 1 ]; do
       tar cf - /home/large/data | ssh remote_host "cat > /dev/null"

   ps. Note that this scenario is common in a large data center, while
       some users transferring large data out of the data center through
       the head node; while other users might copy files and do their
       normal works in the same head node.

4. During the data transfer in the background, I occationally ran this
   command in the same client to test whether there is any abnormality
   in I/O performance (where /home/dir1/file has size 600MB):

   cp /home/dir1/file /home/dir2/

   In the beginning this command can complete in about 1 sec. But after
   around 18 hours (not exactly, because the test ran overnight while
   I was sleeping), the problem appeared. The time to complete the same
   cp command was more than 1 minute.

   During the test, I am sure that the whole cluster was idling. The MDT
   and OST servers did not have other loading. The CPU usage of the testing
   client was below 0.3.

   Then I stopped the test, and let the whole system completely idle. But
   after 3 hours, the I/O abnormality of the same "cp" command was still
   there. Only after I unmounted /home and remounted /home, the abnormality
   of "cp" recovered to normal.

Before and after remounting /home (which I call "reset"), I did the
following tests:

1. Using "top" to check the memory usage:

Before reset:
top - 10:43:15 up 35 days, 52 min,  3 users,  load average: 0.00, 0.00, 0.00
Tasks: 404 total,   1 running, 162 sleeping,   0 stopped,   0 zombie
%Cpu(s):  0.0 us,  0.0 sy,  0.0 ni,100.0 id,  0.0 wa,  0.0 hi,  0.0 si,  0.0 st
KiB Mem : 13232632+total, 13000131+free,   647784 used,  1677220 buff/cache
KiB Swap: 15631240 total, 15631240 free,        0 used. 13076376+avail Mem

After reset:
top - 10:48:02 up 35 days, 57 min,  3 users,  load average: 0.04, 0.01, 0.00
Tasks: 395 total,   1 running, 159 sleeping,   0 stopped,   0 zombie
%Cpu(s):  0.0 us,  0.0 sy,  0.0 ni,100.0 id,  0.0 wa,  0.0 hi,  0.0 si,  0.0 st
KiB Mem : 13232632+total, 12946539+free,   675948 used,  2184976 buff/cache
KiB Swap: 15631240 total, 15631240 free,        0 used. 13073571+avail Mem

   It seems that most of the memory were in "free" state. The amount of
   hidden memory was neglectable. So I did not further investigate the
   amount of slab memory.

2. Using "strace" with the following commands:

   - Before reset (took 1 min of each cp):
     strace -c -o /tmp/log2-err.txt cp /home/dir1/file /home/dir2/

   - After reset (took 1 sec of each cp):
     strace -c -o /tmp/log2-reset.txt cp /home/dir1/file /home/dir2/

   From the log files, the major time consuming was read and write syscalls.
   The others are neglectable.

   % time     seconds  usecs/call     calls    errors syscall
   ------ ----------- ----------- --------- --------- ----------------
   (Before reset)
    71.46    0.278424        1920       145           write
    28.06    0.109322         705       155           read
   (After reset)
    52.92    0.299091        2063       145           write
    46.85    0.264777        1708       155           read

   Before reset, since we have done the cp test for the same file a
   few times, the file was already cached. So the reading time is
   smaller before reset than that after reset (since after reset /home
   was remounted).

   Hence from this result, the time of syscalls looks normal. The
   performance drop seems occuring in other places.

Now I haven't done the investigation of Lustre kernel debug log by enabling
Lustre debug=-1. We will find another chance to do it.

Up to now, any comments or suggestions are very welcome.

Thanks for your help in advance.

Best Regards,


On Thu, Oct 08, 2020 at 01:32:53PM -0600, Andreas Dilger wrote:
> On Oct 8, 2020, at 10:37 AM, Tung-Han Hsieh <thhsieh at twcp1.phys.ntu.edu.tw> wrote:
> > 
> > Dear All,
> > 
> > In the past months, we encountered several times of Lustre I/O abnormally
> > slowing down. It is quite mysterious that there seems no problem on the
> > network hardware, nor the lustre itself since there is no error message
> > at all in MDT/OST/client sides.
> > 
> > Recently we probably found a way to reproduce it, and then have some
> > suspections. We found that if we continuously perform I/O on a client
> > without stop, then after some time threshold (probably more than 24
> > hours), the additional file I/O bandwidth of that client will be shriked
> > dramatically.
> > 
> > Our configuration is the following:
> > - One MDT and one OST server, based on ZFS + Lustre-2.12.4.
> > - The OST is served by a RAID 5 system with 15 SAS hard disks.
> > - Some clients connect to MDT/OST through Infiniband, some through
> >  gigabit ethernet.
> > 
> > Our test was focused on the clients using infiniband, which is described
> > in the following:
> > 
> > We have a huge (several TB) amount of data stored in the Lustre file
> > system to be transferred to outside network. In order not to exhaust
> > the network bandwidth of our institute, we transfer the data with limited
> > bandwidth via the following command:
> > 
> > rsync -av --bwlimit=1000 <data_in_Lustre> <out_side_server>:/<out_side_path>/
> > 
> > That is, the transferring rate is 1 MB per second, which is relatively
> > low. The client read the data from Lustre through infiniband. So during
> > data transmission, presumably there is no problem to do other data I/O
> > on the same client. On average, when copy a 600 MB file from one directory
> > to another directory (both in the same Lustre file system), it took about
> > 1.0 - 2.0 secs, even when the rsync process still working.
> > 
> > But after about 24 hours of continuously sending data via rsync, the
> > additional I/O on the same client was dramatically shrinked. When it happens,
> > it took more than 1 minute to copy a 600 MB from somewhere to another place
> > (both in the same Lustre) while rsync is still running.
> > 
> > Then, we stopped the rsync process, and wait for a while (about one
> > hour). The I/O performance of copying that 600 MB file returns normal.
> > 
> > Based on this observation, we are suspecting that whether there is a
> > hidden QoS mechanism built in Lustre ? When a process occupies the I/O
> > bandwidth for a long time and exceeded some limits, does Lustre automatically
> > shrinked the I/O bandwidth for all processes running in the same client ?
> > 
> > I am not against such QoS design, if it does exist. But the amount of
> > shrinking seems to be too large for infiniband (QDR and above). Then
> > I am further suspecting that whether this is due to that our system is
> > mixed with clients in which some have infiniband but some do not ?
> > 
> > Could anyone help to fix this problem ? Any suggestions will be very
> > appreciated.
> There is no "hidden QOS", unless it is so well hidden that I don't know
> about it.
> You could investigate several different things to isolate the problem:
> - try with a 2.13.56 client to see if the problem is already fixed
> - check if the client is using a lot of CPU when it becomes slow
> - run strace on your copy process to see which syscalls are slow
> - check memory/slab usage
> - enable Lustre debug=-1 and dump the kernel debug log to see where
>   the process is taking a long time to complete a request
> It is definitely possible that there is some kind of problem, since this
> is not a very common workload to be continuously writing to the same file
> descriptor for over a day.  You'll have to do the investigation on your
> system to isolate the source of the problem.
> Cheers, Andreas

More information about the lustre-discuss mailing list