I am wrestling with some perceived performance issues on a 3-node HPE Apollo cluster running Linux. (The cluster supports a backup application.) These are very $$ servers. In our setup each node has 40 CPU's (2 sockets, 10 cores per socket with hyperthreading enabled).

Here is my simple test:

dd if=/dev/zero of=/dev/null bs=256K count=10000

The command reports a 'thruput' of 1.5GB/s. Running the exact same command on my lame HP desktop workstation (Ubuntu) gives 13.7GB/s, and running an another HPE G9 Linux server (not an Apollo) generates 17.8GB/s. This dd command should be virtually all CPU (with a small bit of memory), and on an Apollo server I would have expected a quite stellar 'thruput' measurement for taking an endless stream of zeros and piping them into the bit bucket. But 1.5GB/s?? That seems quite pathetic to me. Am I crazy in thinking something is just not right with the Apollo? What could possibly explain this differential in performance?

  • "command should be virtually all CPU (with a small bit of memory" is that the case? All this command does is take CPU-generated zeros, copy them to memory, and write them to a destination that discards them. Neither generation nor discarding would be time-consuming, but copying zeros to and from the RAM is where the bulk of them time would go. It's easy to check what's the bottleneck - tie it to a CPU core and check if that core gets 100% usage. – muru Jul 10 at 0:44
  • Same results with taskset -c 0 dd if=/dev/zero of=/dev/null bs=256K count=100000 ? – Mark Plotnick Jul 10 at 11:07
  • Does each system have the same kernel version and the same dd? – Mark Plotnick Jul 10 at 11:08
  • Each of the Apollo nodes is identical in kernel version and any executables. Tried the taskset command as suggested and that returned exact same results. Looks like I will be opening support ticket to HPE to get some explanation. – kafir55 Jul 10 at 20:06
  • Do the Apollo and non-Apollo systems have the same versions of dd? – Mark Plotnick Jul 13 at 11:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.