Well, the rule of thumb is that a 10k drive can do 120 to 130 IOps and a 15k drive 170 to 180 IOps.

The improvement of IOps caused by cache depends on your host. If your host does a lot of IOs to the same blocks, the improvement will be enormous, since all IO will be in cache, but as soon as IO has to be done to disks, the cache has to wait for the physical drives.

In the end the cache is there for massive write bursts. And the CX4-120 tot 960 have a certain amount of cache available. The bigger the cache is, the larger these write bursts can be without the hosts noticing slowness caused by disks. The config of the cache can be important as well. Make sure you have far more write cache than read cache as read cache will only be there for sequential reads, while the majority or reads will probably be random. I always assign about 100MB to read cache and then max out the write cache.

Thanks, Is there a thumb rule or a standard way to know how much improvement should be expected?

As RRR said, the System cache improves the performance of the disks - the range given 160 to 180 IOPS (or 10-12MB/s) is WITH cache enabled. As there is no on-disk cache, when you disable cache on the LUN, performance on the disk will be much lower (3-4MB/s).

Overall performance of the disks will depend on the workload from the host - the more sequential the IO the more throughout or bandwidth you can get. I've seen up to 450 IOPS on 15K disks with response times under 20ms - the IO was small block (8KB) highly sequential (streaming) IO.

glen

Since you are asking for a Rule of Thumb, let me throw this out there:

"There is no such thing as cache when designing a RG/LUN Capacity solution."  Always design for the performance you need in the spindle design and do not include the cache in the I/Ops or MB/s calculations.

Hope that helps.

The rule of thumb is 180 IOps for a 15k WITHOUT cache. Furthermore I think I remember reading somewhere that this 180 IOps where 4kB IOs are used. Smaller IOs give more IOps, bigger IOs will produce less IOps.

Cache will of course improve performance, but it simply can't be predicted how much IOps extra you will get from extra cache.