Just wondering if anyone has had a play with Adaptive Flash Cache? I've started it in simulator mode with 128GB as a starting point and I'm seeing an average of just over 2% in additional read hits. At the moment it doesn't seem worth enabling, I think we're better off using the SSD space for AO. Would like to know your thoughts.
If you want to enable AFC in simulator mode the commands are:
The VV set that it's enabled for is 15TB in size so it's not huge, they're enabled for AO as well so most of the hot blocks are probably on SSD already.
Be careful about reaching conclusions based on artificial benchmarks as this probably isn't representative of a real data access pattern.
Being able to gauge the benefits is precisely why the AFC simulator exists, for many Customers especially with smaller active data sets the internal Controller DRAM coupled with wide striping is more than sufficient. It's one of those YMMV, if you need it you'll definitely see big benefits, not only will it accelerate front end I/O but it will also unload the backend disk. For many users, trying to constrain this to a single VVSET may be a mistake, the unloading of the backend (assuming you have enough control memory and SSD) means your probably better off applying this to the system as a whole, after all a rising tide raises all boats.
Also bear in mind that flash cache only works well for certain workloads. It took hp a very long time to enable it, and they kept saying that most of their customers didn't need it.
Even emc recommends that the workload be evaluated before turning on flash cache because for a lot of workloads it does very little
tbh our vSphere data stores and SQL DBs are the only workloads we'd be interested in enabling for AFC and most of their active data has already been punted to SSD by AO. I was surprised at how little difference it's made considering the difference in size between DRAM cache and Flash Cache.... I guess it just goes to show that the majority of I/Os occur in across a small set of data, our cache hits are 88% regardless:
If your data density "hotness" is so small it fits in memory cache, AFC will not improve that much. You get diminishing returns the higher your hit rate...
My understanding is that AFP is quite selective in what to put into cache as well. Only random IO below 64kB?
I wish 3par would have offered cache upgrades over flash cache. I hear the newer V series comes with double the cache the older units came with, but there is no customer upgrade path. I am also interested to test AFC in our DR site as we get killed with performance when we bring up DR since AO has pushed everything down to NL, also eager for write based AFC which I am not sure when that is coming.
I imagine you can buy the DIMMs to upgrade, the issue that I'm aware of is that it can't be done online as node pairs must have the same amount of cache for the redundancy code etc. Hence it would require a shutdown to install. Not sure if the rest of the controller is the same but again if it needed controller swaps it would be a shutdown (not something we expect to do in the life of the arrays ).
There are other considerations as well around pinned data space, hold up power and the internal node SSD disk etc all related to power fail and cache destage to disk etc. So it's not quite as simple as adding more DRAM to a node, but flashcache gives you pretty much the same if not more benefits with a larger capacity and lower cost without any of the potential disruptions. Also keep in mind this isn't the end state for Adaptive Flash Cache, there's a reason it's configured using Raid10 LD's.
).