raw image on NTFS or raw device, performance difference

sunyucong · Sun Apr 29, 2012 7:59 am

Hi,

Can some one with real-world experience talk about the performance difference running starwind using a raw image on an big NTFS file-system vs directly exposing block device? in both cases, it will be on 12 disks raid10

Thanks.

Sun Apr 29, 2012 12:23 pm

No difference. We use NTFS (or any other FS we layer on top of) only to mark extents as "used". All requests are page-aligned (memory), extent-aligned (disk) and don't exploit FS cache (as we use own, much more effective one). It's RECOMMENDED to use IMGs rather then raw devices because of flexibility to move them.

sunyucong · Tue May 01, 2012 8:39 pm

It sounds like you are saying one image for each client , is that right? In term of VMware, Is it better to make one image for each guest and use mapped raw or should use one image per disk array as a vmfs store?

Tue May 01, 2012 10:18 pm

I never said you need LUN-per-VM... Good practice is to store ~20 VMs per LUN. Heavy things like f.e. SQL server require dedicated LUN for optimal performance.

sunyucong wrote:It sounds like you are saying one image for each client , is that right? In term of VMware, Is it better to make one image for each guest and use mapped raw or should use one image per disk array as a vmfs store?

sunyucong · Thu May 10, 2012 11:11 pm

The problem is, I want store about 500 vm on the array, possibly a single lun. And I think should not divide the lun since that will just screws io scheduling in Esxi. what do you think?

Fri May 11, 2012 10:33 am

It's considered to be a very bad practice to store that many VMs on a single LUN. Google a bit about recommended settings but AFAIR it's 20-30 maximum.

sunyucong wrote:The problem is, I want store about 500 vm on the array, possibly a single lun. And I think should not divide the lun since that will just screws io scheduling in Esxi. what do you think?

sunyucong · Sat May 12, 2012 6:56 am

can you enlight me some reasons?
Is this because of command queuing issue or scsi reservation conflict issue?

Sat May 12, 2012 7:04 am

I'm not hypervisor vendor so I can only guess. I think it's because even with hardware accelerated locking the whple thing put too heavy load on a LUN.

sunyucong wrote:can you enlight me some reasons?
Is this because of command queuing issue or scsi reservation conflict issue?