> -----Original Message----- > From: Deneau, Tom [mailto:tom.deneau@xxxxxxx] > Sent: Wednesday, August 26, 2015 5:23 PM > To: Dałek, Piotr; Sage Weil > > > There have been some recent changes to rados bench... Piotr, does > > > this seem like it might be caused by your changes? > > > > Yes. My PR #4690 (https://github.com/ceph/ceph/pull/4690) caused rados > > bench to be fast enough to sometimes run into race condition between > > librados's AIO and objbencher processing. That was fixed in PR #5152 > > (https://github.com/ceph/ceph/pull/5152) which didn't make it into 9.0.3. > > Tom, you can confirm this by inspecting the contents of objects > > questioned (their contents should be perfectly fine and I in line with other > objects). > > In the meantime you can either apply patch from PR #5152 on your own > > or use - -no-verify. > > Piotr -- > > Thank you. Yes, when I looked at the contents of the objects they always > looked correct. And yes a single object would sometimes report an error and > sometimes not. So a race condition makes sense. > > A couple of questions: > > * Why would I not see this behavior using the pre-built 9.0.3 binaries > that get installed using "ceph-deploy install --dev v9.0.3"? I would assume > this is built from the same sources as the 9.0.3 tarball. No idea actually. It's a race condition, so it might be just luck. > * So I assume one should not compare pre 9.0.3 rados bench numbers with > 9.0.3 and after? > The pull request https://github.com/ceph/ceph/pull/4690 did not mention > the > effect on final bandwidth numbers, did you notice what that effect was? That depends on the CPU performance, but you should expect differences of a few or tens of MB/s on smaller blocks to up to even hundreds of MB per second on larger block sizes. More concurrent jobs also make issue more visible and add to total difference. With best regards / Pozdrawiam Piotr Dałek _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |