2

My understanding is high kernel/sy CPU usage is a symptom of network and disk IO or issues of RAM throughput. #516139. However, I suspect in the case below, over allocating threads is giving the kernel (via scheduling) way too much to do and the actual user level process computations are suffering.

We've parallelized building many models in R without realizing that each model building function is openMP capable and will default to distributing itself (to all available cores!?).

  1. Without already having suspicious/poorly written code to reason about, is there a way to tell sy usage is high because of thread allocation?
  2. Once this is running is there anyway to to set e.g. ulimit on an individual process or any other recourse short of killing the top level process?

htop screenshot

mpstat

Linux 4.9.0-4-amd64 (rhea.wpic.upmc.edu)        01/19/2018      _x86_64_        (72 CPU)

11:27:42 AM  CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
11:27:42 AM  all   13.28    0.00   30.09    0.17    0.00    0.03    0.00    0.00    0.00   56.42

iostat

Linux 4.9.0-4-amd64 (rhea.wpic.upmc.edu)        01/19/2018      _x86_64_        (72 CPU)
avg-cpu:  %user   %nice %system %iowait  %steal   %idle
          13.28    0.00   30.16    0.17    0.00   56.40

Device:            tps    kB_read/s    kB_wrtn/s    kB_read    kB_wrtn
sda               5.90       182.64       532.33  175621086  511871712
sdb               4.26        10.68      1014.29   10268992  975304572
sdc               0.82        14.68        20.13   14111683   19354860
sdd               0.00         0.02         0.00      18100          0

cat /proc/self/mountstats

device skynet:/Volumes/Phillips/ mounted on /Volumes/Phillips with fstype nfs statvers=1.1
        opts:   rw,vers=3,rsize=65536,wsize=65536,namlen=255,acregmin=3,acregmax=60,acdirmin=30,acdirmax=60,hard,proto=tcp,timeo=600,retrans=2,sec=sys,mountaddr=xxxxxxx,mountvers=3,mountport=748,mountproto=udp,local_lock=none
        age:    960031
        caps:   caps=0x3fc7,wtmult=512,dtsize=32768,bsize=0,namlen=255
        sec:    flavor=1,pseudoflavor=1
        events: 22290816 111452957 1493156 351729 8727855 16583649 130526167 54024016 266 1063322 0 8965212 14276120 4723 2406702 480455 238 1439836 0 615 53950807 7 0 0 0 0 0 
        bytes:  1584954051456 218684872379 0 0 742185394287 219176285117 181264042 53636171 
        RPC iostats version: 1.0  p/v: 100003/3 (nfs)
        xprt:   tcp 1017 1 75 0 0 66894351 66887373 6904 256266328938 0 802 1887233163 595159288
        per-op statistics
                NULL: 0 0 0 0 0 0 0 0
             GETATTR: 22290802 22290914 0 3154213512 2496568872 18446744073371231314 88830744 118185897
             SETATTR: 5616 5618 0 942564 808704 122600 893047 1025591
              LOOKUP: 16586987 16586993 0 3230313244 3836903464 18446744073421412542 29327650 31652035
              ACCESS: 5630423 5630439 0 810455208 675650520 2233531 21149691 23526686
            READLINK: 60834 60834 0 9245324 9267896 269 957051 958788
                READ: 11461667 11461844 0 1688228580 743652637248 160174235 1277881121 1438304121
               WRITE: 4246754 4259238 220 220002658844 679480640 30785630990 5061286597 35853150454
              CREATE: 7464 7467 0 1485604 1970496 801177 746707 1551420
               MKDIR: 83 83 0 16296 21912 1749 1164 2986
             SYMLINK: 30 30 0 8504 7920 0 16 34
               MKNOD: 0 0 0 0 0 0 0 0
              REMOVE: 9276 9278 0 1742408 1335744 143237 439704 583661
               RMDIR: 78 78 0 13080 11232 0 68 78
              RENAME: 908 908 0 214236 236080 2906 27182 30095
                LINK: 0 0 0 0 0 0 0 0
             READDIR: 204340 204340 0 32694564 6032970656 42323 1722666 1771971
         READDIRPLUS: 6343408 6343410 0 1040350176 31022488528 1465418 136921691 138608729
              FSSTAT: 2834 2834 0 388096 476112 67600 532404 600234
              FSINFO: 2 2 0 224 328 0 0 0
            PATHCONF: 1 1 0 112 140 0 0 0
              COMMIT: 35880 35964 1 5029968 5453760 41064204 31974116 73123499

edit:

This situation exists because openblas defaults to parallel computations. see

1

At a glance, 400 load average seems high for a 72 CPU core box. More tasks ready to run than cores usually means some of those are waiting.

System time can be a number of things. For compute bound workloads like what yours might be, 30% system CPU seems high.

To see exactly what is happening, you can sample call graphs system wide, then make them into a neat visualizations called flame graphs.

# Looks like you have a Debian install
# Install debug symbols for the kernel and Linux perf
apt-get install linux-image-amd64-dbg linux-tools
git clone https://github.com/brendangregg/FlameGraph  # or download it from github
cd FlameGraph
perf record -F 99 -a -g -- sleep 60
perf script | ./stackcollapse-perf.pl > out.perf-folded
./flamegraph.pl out.perf-folded > perf-kernel.svg

The widest plateaus in the graph should indicate where the most time is being spent.

What to do about this depends on what you find. I'm not familiar with openMP, but if it indeed it already runs in parallel, limit the number of concurrent jobs. Don't have them fight each other for resources.

800% CPU on a task does imply that you have multi threaded tasks, using perhaps 8 cores. If this is typical, running 8 or 9 of these will keep 72 cores utilized. There are ways to run scripts in parallel until a certain load level is reached, in particular GNU parallel.

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.