Now we are doing final testing with bigger SSD RAID array, with 25% of space left unprovisioned, and using only ~30% of the capacity to tasks which most require that performance boost - and it's still not looking worth the effort and cost. Infact, the only scenarios i've seen SSDs actually perform anywhere the promise land are single drive and RAID0, never as part of a cache, but as sole storage.
Therefore, we have purchased a set of 146Gb SAS 15k drives for testing purposes - if these provide the kind of performance boost we are looking for in a RAID10 setup. Ironically, at best you can get 6x SAS 146Gb drives for the price of 1 SSD drive. 15k performs at roughly 350 IOPS no matter what - stable performance like all magnetic drives. That means, with proper setup we gain about 2100 IOPS on the price of 1 SSD drive - which we have not seen to provide more than 500 IOPS in production, at the cost of higher electrical consumption and less ports for the big storage.
16x 7200RPM provides about 1920IOPS, but in RAID50 setting (8xRAID5 + 8xRAID5 in RAID0) you get approximately 85% of that leaving you with roughly 1600 IOPS *max*, in practice more like 1500 IOPS. So in a chassis of 24 drives, adding 8x SAS 15k will add 2800IOPS, in real world probably 2500 IOPS, almost tripling the performance, of course, only when properly setup.
We will see how that goes, we get to do some preliminary tests in end of November when sea freight container arrives with new storage chassis's.
יום שני, אוקטובר 28, 2013
