HPE Storage Users Group

Tracking cache pages on any flashcache implementation would require metadata and on a 3PAR metadata sits in control memory, everyone else's that I'm aware of sits in shared data cache. Since a smaller array has less control memory it would make sense that the smaller array would support less flash cache. Either that or like the other implementations out there you lose DRAM speed data cache in order to support NAND speed flashcache. But like all read caches in a random environment the bigger they are, the less efficient they become. So having a massive read cache is not necessarily a good thing, or at least not a good use of relatively expensive resource.

A write cache on the other hand can be very efficient but requires much higher levels of protection and testing and at some point you still need to destage to disk to ensure consistency and that is going to take an inordinately long time for a 10TB cache. I have yet to see anyone implementing anything like those kinds of capacities for a write cache as it doesn't really make much sense. They're there to support transient spikes not your entire workload and if you don't have a backend that can destage quickly enough, then you'll be in a world of hurt if the cache ever fills, or indeed in certain failure scenarios.

With the advent of inline dedupe and low cost SSD, if you choose you now have the potential to just bypass a SSD based write or read cache completely and go straight to disk. After all it's the same SSD's, reading and writing at the same speed and if you can put 10TB's on the floor with the addition of 4:1 dedupe on top then maybe flashcache no longer holds the same importance it did for spinning disk. Not saying it won't be extremely useful but there are now potentially other commercially viable ways to skin that particular cat.

Timing of product releases isn't anywhere near as simple as a roadmap might suggest, HP aren't just looking to release feature the moment they are ready, they're looking to release features that are complimentary and provide synergies. That way Customers don't go off at a tangent implementing a particular feature only to then have to roll it back when the next comes along. The aim being to get away from those mutually exclusive features that plague much of the competition. BTW That's the No1 rule of all roadmaps - dates are very liable to change

cleanur,

I agree with your points, but I think we need to look at what flash cache is trying to solve. I have no inside info as to what the feature is targeted at solving, but to me it should be solve the problem with the length of time it takes data to move through AO. The biggest issue we see is data that is stale becomes hot and now has to wait before it gets moved back to a faster tier. To me having a a "cache" area to handle these blocks that resurrect from the dead until AO can move them is important. To limit that means I still have a potential bottle neck if too much data becomes hot and the "flash cache" area is not large enough. We are constantly struggling with AO dropping data to bottom tier and not utilizing the faster tiers enough and then data comes to life and has to wait. This is very evident when we do DR testing as our DR array has almost all of its data falling to lower tiers since it is inactive data.

As I have mentioned in other threads AO is a mixed blessing and I would never use as much NL drives again (if at all). Our new 7400 is SSD and 10K only and it screams. Our V400 is 15K, 10K and NL and the NL drives are being crushed.

As expected Dell was very aggressive on pricing, $2/GB with 20% flash storage. Quote includes 5 years hardware and software support. I think we're going to pull the trigger, I'll let you guys know how it compares to our 7400-4.

Look out for those 4TB drives

I'm looking at a similar low $ per GB Compellent configuration in which Dell are seriously proposing over 85% of usable capacity be provided on 4TB 7200 RPM Nearline drives running Raid 6 (and this isn't the first). Not a good place to be if you're expecting any level of consistent performance beyond the first few days of testing.

As per hdtvguy and AO tiering, a very large flashcache would definitely be of help in this kind of configuration but in reality it would be needed to overcome a flawed design, rather than to enhance performance.

I own a pair of Compellents and 7 3par arrasy and I will take any 3par over any Compellent any day of the week. The Comepllent is decent, and there are features they have I wish 3par had, but Compellent is just not up to the task.

Some basic differences that we value:

- 3par truly leverages all controllers for all volumes, Compellent, each volume is owned by a controller and IO for a volume only goes through that controller.

- Compellent is only a dual controller setup, 3par you can do 2,4 or 8 controllers.

- 3par has a very robust CLI that you can virtually script everything and anything through, Compellent does not.

- To fully leverage some of the Compellent features you need to have their Enterprise Manager software running on system.

- 3par has true IP based replication, Compellent is either FC or iSCSI

Why would I be any more scared of 4TB drives in the Dell than I am of the 3TB drives in my 3Par? High TB/Watt for stale data was a design requirement, not something I fear.

One word ------ Rebuilds

But if it's for stale data, then I could think of plenty of even more economical places for the data to reside. Don't want to preach to the choir but If you're seeing a big difference in pricing in the midrange, especially on equivalent platforms then someone has the wrong end of the stick.

Rebuilds are an issue with drives of that size for sure.

The other issue specific to 4TB drives is the IO density. The tendency for some vendors is to use cheap disk to meet capacity requirements, but it doesn't meet the performance requirements. Now these are NL drives so performance requirements should be pretty low, but with data density that high, the tier will be a pig. The other thing to remember is that a 4TB drive will perform slower when compared even to a 2TB or 3TB drive.

Just make sure that for each tier, both your performance AND capacity requirements are being met.

It's easy to meet the capacity requirements, forget about performance, and come in low.

I have a Compellent in the lab as well. I recall something funky about the zoning/multipath setup... active/passive, no round robin, and a license key required to enable/convert host ports to use NPIV for failover instead of physically dedicating a host ports as a standby for partner node failover.

_________________
Richard Siemers
The views and opinions expressed are my own and do not necessarily reflect those of my employer.

The Compellents also need their physical ports added to zone to themselves because that is how the controllers communicate over the FC network. So you wind up with a zone on each fabric that contains the physical controller ports and then zones with the controller virtual WWNs to the respective hosts. Very Mickey Mouse to me. I do not consider Compellent in the same league as 3par.