Christian Rohmann put forth on 11/3/2010 10:02 AM: > Maybe any1 has more ideas based on the fact that the > thing is stable with two cores now, but wasn't with eight.
Absolutely. With 8 virtual CPUs (gasp OMG! big no-no) your guest kernel will be generating a vastly larger number of timer interrupts/second that the VMkernel has to service. I believe this is covered in the VMware article Wietse referred you to. This is one of the reasons VMware recommends you only configure a VM guest with the minimum number of vCPUs needed, which is almost always ONE vCPU. VMware is a consolidation platform, not a performance platform. The only reason ESX was given the ability to use virtual SMP is for customers who deploy massive databases and such atop ESX in a SAN environment strictly for 2 main reasons: 1. VMotion HA 2. Consolidated backup VMware did not envision customers doing consolidation _and_ giving each guest a high count of vCPUs. I've seen cases studies of real environments running huge Oracle databases on 8 socket x86 servers with 128GB RAM, with ESX underneath providing 8 vCPUs. This is why the feature was added. It wasn't added so Joe admin could run his 20 LAMP servers with 8 vCPUs each. You're not the first OP I've seen or heard of getting in trouble with this. Unfortunately EMC is all about making money. One usually only learns of this when going through paid training. Another avenue is asking the right questions of one's vendor ESX engineer who has gone through the training. One other avenue to this information is talking up a VMware company tech at a trade show/conference. This is where I learned of this vCPUs issue, some 4 years ago. Obviously you can learn of it on the Postfix mailing list as well. ;) -- Stan
