Performance Difference Between ESX and ESXi
Has anyone seen a performance difference between the two version of ESX? I have a dell R905 running ESXi on a embedded usb. I was wondering if it would run better if on loaded ESX on local 15k drives instead.
I have been doing some bench marks with the R905 with native windows, compared to an R905 with exsi4 and a 2003 vm on top of it and the vm gets about 1/3 the native disk performance. Not sure what the problem is. THe are both connected to the same Fiber channel SAN
I have been doing some bench marks with the R905 with native windows, compared to an R905 with exsi4 and a 2003 vm on top of it and the vm gets about 1/3 the native disk performance. Not sure what the problem is. THe are both connected to the same Fiber channel SAN
ASKER
I thought the since the vmkernal swap file is on internal usb, rather than a fast hard drive might affect performance.
I used iometer. tested sequental read, 2k blocks.
Testing to see if it makes sense to put a faily large SQL 2005 db on a vm.
I used iometer. tested sequental read, 2k blocks.
Testing to see if it makes sense to put a faily large SQL 2005 db on a vm.
SQL server will normally have good performance on VMware, especially on RVI/EPT systems (83xx Opteron / 55xx Xeon).
The vmkernel swap files are per default stored in the same location as the VM's, but it can be configured to use local storage. Even though these files exist, they will not have any activity unless you're using >~80% ram on your system.
Maximum sequential read would be higher on a native server, but an sql workload is normally not such a workload.
I have also seen RAID10 give much better performance on LUNs hosting SQL server than RAID5 or 6.
Lars
The vmkernel swap files are per default stored in the same location as the VM's, but it can be configured to use local storage. Even though these files exist, they will not have any activity unless you're using >~80% ram on your system.
Maximum sequential read would be higher on a native server, but an sql workload is normally not such a workload.
I have also seen RAID10 give much better performance on LUNs hosting SQL server than RAID5 or 6.
Lars
Here are some documents on running SQL Server in VMware:
SQL Server Performance in a VMware Infrastructure 3 Environment
Best Practices: Microsoft SQL Server and VMware Virtual Infrastructure
Lars
SQL Server Performance in a VMware Infrastructure 3 Environment
Best Practices: Microsoft SQL Server and VMware Virtual Infrastructure
Lars
ASKER
Thanks for the replies, I have read that doc and a few others, we are doing everything that is recommend. Our SAN vendor(compellent) also has a best practice doc and we are to spec. The test volumes i used were on RAID10. We have 59 Disk in the system.
Phyical server was getting about 5000IOPS
Virtual server was getting about 1600IOPS.
There is a bottleneck some where, i just need to find it. So far it is pointing the the vmware software.
Both servers are Dell r905's 4 Quad core AMD's 8384, 128GB RAM, 8 NIC, 4-4gb fiber hba's. running vsphere 4 enterprise plus
Test was windows 2003 enterprise loaded natively on the server with one connected to a 100GB Raid 10 Lun, Other was a windows 2003 enterprise loaded on the esx server, with a phsyical RDM 100GB Raid10 LUN
Phyical server was getting about 5000IOPS
Virtual server was getting about 1600IOPS.
There is a bottleneck some where, i just need to find it. So far it is pointing the the vmware software.
Both servers are Dell r905's 4 Quad core AMD's 8384, 128GB RAM, 8 NIC, 4-4gb fiber hba's. running vsphere 4 enterprise plus
Test was windows 2003 enterprise loaded natively on the server with one connected to a 100GB Raid 10 Lun, Other was a windows 2003 enterprise loaded on the esx server, with a phsyical RDM 100GB Raid10 LUN
ASKER
I have opened a call with my SAN vendor. Will report back
ASKER CERTIFIED SOLUTION
membership
This solution is only available to members.
To access this solution, you must be a member of Experts Exchange.
That is great news! Good you figured out these issues.
There is some overhead by running in VMware, but the random IO should be very similar to running a native system while sequential IO has a larger gap. Random IO is the most important thing for most systems.
How did you measure your performance?
Lars