We had an initial call open with VMware with performance between ESX and a HP MSA iSCSI device. The client had a SQL application and were running a specific report to benchmark the performance. The iSCSI performance vs the direct attach physical counter part lagged way behind. We also used a little third party tool called atto that showed that the throughput was only 140MB
VMware suggested that the cause was down to an I/O bottle neck on the MSA. We have now replaced the iSCSI MSA with a Fibre Channel P2000 MSA and installed 8Gb FC cards in the ESX servers.
The benchmark tool that we used which showed 120MB throughput on the iSCSI is now showing nearly 700MB on the Fibre. So we feel that the bottle neck should be gone.
However the SQL application within the virtual machine, although slightly faster than when it when it was connected to the iSCSI LUN, is still slower than it's physical counterpart. The SQL virtual server has been split across drives (LOGS, 3 disk raid 5 and data a 8 disk raid 10) using paravirtual drivers to attach to the vdmk files. The paths to the LUNs has been setup using MRU and Round Robin with no effect on performance. In fact any change I make to the MSA in regards to disk cache or write optimisation makes no difference to the performance of the virtual machine, leading me to think that the perfomance issue might be in our setup of VMware. We have explained the purpose of VMware is not for performance improvements, but the client is now disappointed as they feel that with the added spindles and faster backbone the server shouldn't just match the current setup but it should absolutly fly!
We are using the latest build of vSphere 4.1. Does anyone have any suggestions about tweaks or possible perfomance improvements that we might have missed.
