According to what Texiwill and some posts I read, this is a very finicky protocol. You get a SCSI reservation any time a file is opened or closed. This includes, ISO images on VMFS and various other files (like logs, which is why I turn those off for each VM).

We still get them no matter how much qladepth I give, or tweak, or any type of testing on the SAN, nothing seems to change the fact that SCSI reservations persist. Interesting thing is the 'threshold' you seek happens on 1 VM. I setup an ESX server, and purposely tried to repeat the issues hoping to figure what could cause such a problem. I had all kinds of monitors on Fiber Switches, the cards, I saved logs, I did everything. I called VM Ware, QLogic and no one has a clue. So I gave up.

Another odd thing, QLogic spits out WAY more messages than Emulex. So this MUST be a driver issue, not ESX / SAN connectivity. Performance doesn't suffer unless the depth is too low, but when we got IBM 3950 with emulex (keep in mind SAN, VM's, Fiber, Switches nothing changed...) the SCSI reservation is almost gone completely.

Interesting. By comparison Dell 2950 (not picking on Dell, but it's the majority of machines we have) almost have 20 or 30 SCSI reservations an hour, which is the same machine I used to test just 1 VM, even with 1 VM, it still had a SCSI reservation.... during power on, off, and access (something that perplexed QLogic, because they claim that should not happen). But we have Dell R900 with Emulex, and SCSI reservation were cut in half but still numerous. Only the IBM machines show virtually no SCSI reservation, even using default qladepth at 16, and they have almost 4 times the VM's.

So if any machine would show SCSI reservation, you would think the ESX server with the most VM's would have more SCSI reservations, but they don't. I conclude there is a bottleneck somewhere on the backplane, which explains why IBM have better Disk IO than the Dell counterparts on the SAME SAN / Fiber storage.

So I think this is more of a combination of things, driver, ESX version and hardware baseline. The best combination thus far (In my opinion is IBM, not knocking Dell, I am a huge Dell fan) but numbers speak for themselves.

In the end, the R900 have no performance degradation on IO, again the difference is QLogic vs Emulex. bottom line I think is Emulex is a better hardware / driver combo than QLogic.

Hello,

There is no real 'great' number. THe problem is that if you get a SCSI RESERVATION message it does not mean there is a conflict. Just means there was a reservation request made and the SCSI commands had to be redone but no conflict occurred. When the CONFLICT does occur there is a chance that data was not written to disk.

CONFLICTS are the bad things. If you get those regularly then something else is happening, but reissues are quite common. Also, if you get many conflicts the guest OS should pick up on this and reissue the request as well (depending on the guest OS of course).

So if you get SCSI RESERVATION messages, those are fine. If you get CONFLICTS that is something entirely different and if you get a lot of them you may want to talk to the SAN vendor..... Hitachi had a firmware issue at one point.

VMware ESX has gotten much more resilient to SCSI/FC issues with respect to conflicts.


Best regards,
Edward L. Haletky
VMware Communities User Moderator, VMware vExpert 2009, DABCC Analyst
====
Author of the books 'VMware vSphere(TM) and Virtual Infrastructure Security: Securing ESX and the Virtual Environment' available for pre-order now
'VMWare ESX Server in the Enterprise: Planning and Securing Virtualization Servers', Copyright 2008 Pearson Education.
SearchVMware Pro|Blue Gears|Top Virtualization Security Links|Virtualization Security Round Table Podcast