openstack team mailing list archive

[Jay Pipes <jaypipes@xxxxxxxxx>]

Hi all,

If you're interested in helping me gather throughput numbers for
various Glance installations, please contact me. I wrote a little tool
tonight that gathers some throughput details after attempting to
concurrently add images to a Glance server.

You can see the output of the tool below:

jpipes@librebox:~/repos/tempest/tempest$ ./tools/glance-perf -v
--processes=4 --create-num=12 --create-size-min=20
Running ./tools/glance-perf against librebox
 PROCESSES: 4
 WORKLOAD: Create 12 images with size between 20MB and 100MB
 CREATING LOCAL IMAGE FILES (this may take a while)
  Creating local image file /tmp/tmpKVyclj/0.img of size 25MB ....... OK
  Creating local image file /tmp/tmpKVyclj/1.img of size 56MB ....... OK
  Creating local image file /tmp/tmpKVyclj/2.img of size 54MB ....... OK
  Creating local image file /tmp/tmpKVyclj/3.img of size 53MB ....... OK
  Creating local image file /tmp/tmpKVyclj/4.img of size 54MB ....... OK
  Creating local image file /tmp/tmpKVyclj/5.img of size 72MB ....... OK
  Creating local image file /tmp/tmpKVyclj/6.img of size 20MB ....... OK
  Creating local image file /tmp/tmpKVyclj/7.img of size 56MB ....... OK
  Creating local image file /tmp/tmpKVyclj/8.img of size 65MB ....... OK
  Creating local image file /tmp/tmpKVyclj/9.img of size 74MB ....... OK
  Creating local image file /tmp/tmpKVyclj/10.img of size 80MB ....... OK
  Creating local image file /tmp/tmpKVyclj/11.img of size 58MB ....... OK
 DONE CREATING IMAGE FILES
 CREATING WORKER POOL
 CREATING IMAGES VIA WORKERS
Total time to add images: 7.63540 sec
              throughput: 87.35624 MB/sec
 REMOVING LOCAL IMAGE FILES
 REMOVING IMAGES FROM GLANCE

Anyway, if you're interested, I'm looking for feedback on a number of things:

a) The types of workloads to add to the tool (currently tests
concurrent uploads of images, but obviously I want to test real-world
read/write workloads and am interested in getting feedback on what
kind of read/write ratios you see in production (# index vs. # add vs.
# delete, etc)
b) I've built the tool to create image files using output from
/dev/urandom of a configurable size (shown above is random image size
between 20M and 100M). I'd like to get some feedback on the size of
real-world images in use in production systems and then see what the
impact of image size is to throughput
c) If you have test systems that you wouldn't mind running my little
tool against, I'm interested in gathering feedback on the throughput
of the various Glance backend drivers -- Swift, S3, filesystem, RADOS,
etc

Lemme know by pinging me on email.

Cheers,
-jay