CRAZY swings in datastore access times

ofc2logic · Sun Feb 17, 2013 12:23 am

hello.

I have StarWind iSCSI SAN running on a 2U server with six 1TB SATA drives. The cache set on the 2U server is equal to the hardware- based memory cache on the PERC6/i RAID controller.

Take a look at the attachment. You'll notice that there are periods when access times are STELLAR, and then periods when it swings WILDLY in to the 300-800ms range.

The virtual machine is quite SNAPPY when read and write access times are 1ms, and quite SLOW when access times suddenly peak.

Is there any logic to this? I'm dumbfounded.

Sun Feb 17, 2013 12:29 am

It's a bad idea to do what you're doing: trying to match StarWind's cache to the size of hardware RAID controller and using write back cache on a single node. Slow and at the same time dangerous. Wise would be allocating ~1/100th of a served capacity (say 10 GB of RAM for each 1 TB of disk) to be used for so-called "spoofing" and using dual or better triple controller setup with distributed write back caches kept coherent.

ofc2logic · Sun Feb 17, 2013 12:32 am

ok, good to know...

my server has 8GB of memory and ONE RAID controller.

The image file is 3.7TB, though approximately 600GB of that is actually used.

Shall I enable write-through caching with 4GB cache? Is that an effective alternative?

Sun Feb 17, 2013 12:39 am

Does not matter how many RAIDs do you use, it does matter how many hosts running hypervisors you use.

Yes, going with 4GB WT (WB for experiments) should be fine for now.

ofc2logic wrote:ok, good to know...

my server has 8GB of memory and ONE RAID controller.

The image file is 3.7TB, though approximately 600GB of that is actually used.

Shall I enable write-through caching with 4GB cache? Is that an effective alternative?