In this article I want to set the record straight regarding last week’s article about New in esxtop RESV/s (SCSI reservations per second). During VMware Communities Roundtable Podcast number 118 with Krishna Raj Raja. I have learned that I didn’t measure the right metric with the correct method so now I’ve used the right method and metric -> CONS/s. I hope you will enjoy it. You can also watch the HD version of this video on <YouTube> and <ScreenCast>.
I’ve received an email from Krishna Raj Raja with some additional details about SCSI RESV/s and CONS/s counters.
- A nonzero value for RESV/s counter indicates that the host is doing some metadata operation on the VMFS volume (like creating a file, allocating disk space, file rename, permission change, timestamp update etc) - this by itself does not cause any performance problem.
- SCSI reservation conflict counter - CONS/s will become non-zero when a host tries to do any I/O on a LUN which has a SCSI reservation in progress.
- SCSI reservation is held for a very short period (few hundred microseconds) so the chances of a getting a conflict is very less on a small cluster. However as the number of hosts that shares the LUN increases conflicts could arise more frequently.
- SCSI reserve/release commands have higher priority and they are placed ahead in the host's SCSI I/O queue for immediate dispatch.
- When SCSI reservation is held on a LUN, I/O issued by a VM going to that LUN will fail with the BUSY status within the guest, and the guest will have to retry the I/O. This could result in an end to end increase in I/O latency for that particular I/O.
- The new vSphere 4.1 feature - VAAI - has support for ATS (Atomic test and set) functionality which is a replacement for SCSI reservation and it provides much faster/granular locking mechanism. Performance impact is very little if ATS is used instead of SCSI reserve/release mechanism.