Speakers:
Julien Collet
(CERN)
,
Roberto Valverde Cameselle
(CERN)
Giuliano
- CEPH-993: rgw_max_concurrent_requests:
- Set up a VM (warp-xl-benchmark) to issue concurrent get requests to one of the nethub cluster rgw
- For each value of --concurrent ranging between 500 and 1500, the option was tuned to the following values {1024,512,256,128}
- What is observed is that, indeed, each time to you divide by two this number, you roughly divide by two the RSS (except 512/1024 that show roughly same memory footprint,to be investigated? Might be the VM that cannot really issue all these requests at the same time.)
- One could relatively clearly see that the throttling was kicking in since warp was reporting: "warp: <ERROR> download error: Error response code SlowDown."
- TL; DR
- I think it might pay off to tune it to 512 and see if this needs further tuning afterwards.