For FlashPool you can set caching policy per volume:
• none - Does not cache any user data or metadata blocks. • auto - Read caches all metadata and randomly read user data blocks and write caches all randomly overwritten user data blocks. • meta - Read caches only metadata blocks. • random_read - Read caches all metadata and randomly read user data blocks. • random_read_write - Read caches all metadata, randomly read and randomly written user data blocks. • meta-random_write - Read caches all metadata and write caches randomly overwritten user data blocks. • noread-random_write - Write caches all randomly overwritten user data blocks. It does not do any read caching. • random_read_write-random-write - Read caches all metadata, randomly read and randomly written user data blocks. It also write caches randomly overwritten user data blocks.
-----Original Message----- From: toasters-bounces@teaparty.net [mailto:toasters-bounces@teaparty.net] On Behalf Of Michael Bergman Sent: Tuesday, June 30, 2015 5:00 PM To: Toasters Subject: Re: FAS2xxx Upgrade - Flash Cache Performance Estimation?
Stephen Stocke wrote:
We're looking to upgrade two sites which consist of a FAS2020 (7.3.7) and FAS2240-2 (8.1.4-7m) at each site. The 2240-2 systems have a FlashPool accelerated aggregate and the customer is happy with the performance.
So these small 2240-2 boxes, have how much FlashPool each...? One Aggr I assume?
The new filers will be either a FAS2552 with additional shelves or we may ditch the entry toasters and go with a FAS8020 system and shelves at each site. If we go with the 2552 we’ll keep using FlashPool. If we go with the 8020 we have the option of going FlashCache.
So your choice is A) a new FAS2252 with add shelves and FlashPool, or B) a new FAS8020 with FlashPool or FlashCache
is that it? If everyone's good with how FlashPool works in the older/smaller 2240-2 systems you kind of have a basline I'd say. Why would you want to switch over to FlashCache at all? It won't improve anything, and it's not cheaper. The only difference before 8.3.1 (which is out in RC1 now) is that w FlashPool you cannot tune things quite as much as you can in FlashCache since a very long time now (lopri_blocks, large seq R benefits from this):
$ naXXX options flexscale flexscale.enable on (same value in local+partner recommended) flexscale.lopri_blocks on (same value in local+partner recommended) flexscale.normal_data_blocks on (same value in local+partner recommended) flexscale.readahead_blocks off (same value in local+partner recommended) flexscale.rewarm on (same value in local+partner recommended)
It's really not trivial to say neither how much cache you should have, nor if you're better off with lopri_blocks off or on. It is so very dependend on the workload parameters (temporal locality and working set size as well as the mix of ops: random/seq). Personally I feel that it's much easier to understand how FlashCache works, than what FlashPool is doing. Very very different -- vastly dissimmilar places in the whole system so to speak.
For FlashCache you watch the "churn" through the cache and then it's just the Hit Rate and Disk Reads Replaced that matters (higher = better)
Your customers are happy with the performance of the entry level (tiny, from my PoV) 2240 and 2252, so I think this is sub-optimisation at a higher level ;-)
Stephen Stocke wrote:
We'd like to evaluate the potential performance improvements of using FlashCache on the 8020 in order to help with the purchasing decision.
There's likely no difference for you in an 8020 between FlashPool and FlashCache. Especially with 8.3.1 where the -Pool is more tunable than before [see above]
I'd just continue to use FlashPool since you're already there now. No point in starting to use FlashCache too, -Pool is the way fwd in any case and an FAS8020 can have more and more flexible configs of -Pool than -Cache, so...
/M
_______________________________________________ Toasters mailing list Toasters@teaparty.net http://www.teaparty.net/mailman/listinfo/toasters