Both Disk and Cache

Steve Foskett on his blog at PackRat asked the question recently; Is Flash a Disk or a Cache? and listed various vendors and the group they fell in.

I couldn’t answer him at the time, much as I wanted to, but now I can. NetApp isn’t just in the one camp, it’s in both.

Flash as Tier 0

Just about everyone has announced some form of flash drives, mainly using them for tier 0 storage. (Question: when the industry comes up with something even faster and bigger, say holograms, are we going to have tier –1?).

There’s always a workload that can benefit from faster disks. But by partitioning  storage into so many tiers, that introduces a problem; you’re always playing off performance vs capacity and cost. Tier 0 is small and fast, and very expensive.

Flash as Read Cache

With SSD as disk, you’re still stuck with tier 1, 2 and so on. Big and slow, but so much cheaper for the volume they can carry.

A few vendors have indicated that they’re going to use SSD flash as read cache. That has the benefit of making all storage go faster; hot data gets cached in memory that is considerably cheaper than DRAM, and is a lot faster than doing I/O to disk, especially slower SATA devices.

With SSD flash as cache, Tier 1 and above gain from the speed that SSD flash provides.

NetApp: Disk and Cache

The NetApp roadmap includes both versions; flash as disk, and flash as read cache.

Best of both worlds.

[Updated 13Nov: a sharp eyed reader spotted the difference between flash as memory and flash organized as disk – SSD. I’ve corrected the mistakes.]

.

Comments

[Stephen Foskett]

Thanks for the update, Alex! I love the flash read cache idea - it really leverages NetApp's architecture and ought to have some serious benefit. But will it emphasize the existing imbalance between read and write performance?

[Martin G]

Of course SSDs can be used for both and the boundaries will become more and more blurred. Automatation of storage tiering at the block level will be the big story I suspect over the next eighteen months; this automation is absolutely key to sweating the flash asset.

And as for holographic storage, we're probably into the world of imaginary numbers.

[Alex McDonald]

Stephen

Good point. There's always a bottleneck. SSD is going to show up the bottlenecks in everyone's technology. But from a NetApp perspective, we’re better placed than most.

We typically respond extremely quickly to writes because we are able to commit them as they are logged to NVRAM (non-volatile RAM) and don't have to wait for them to go to disk. This makes us fast at writes. When we do have to go to disk before we can respond, we write in such a way that we're always taking as much advantage as we can of the open areas in our layout. This keeps us fast at writes even under heavy load. Kostadis covered some of this in his blog

Unless we already have the data in our system memory, a read will very often require a disk operation to satisfy it. Because system memory isn't as big as the disk that backs it, we often have to go all the way to disk and reads can be slower than writes in some workloads. We currently mitigate this with sophisticated readahead algorithms. I recently covered some of the readset technology in a recent blog entry.

Our caching architecture means that things you've recently read, and are likely to read again soon, will be sitting in a fast location like RAM in current PAM (our Performance Accelerator Module) or in flash in future versions, rather than on disk.

Both our writes, which are already very fast and efficient, and our reads will be highly optimized and the difference between a workload's write performance and read performance will be greatly reduced.

Thanks for commenting.

[Basavaraju]

Hi

1.Netapp PAM (Pisces) card uses flash memory or DRAM DIMM (ECC supported) in their PAM module? Because i see 4 DIMM 4GB each in one card.
2.Is there any new generation PAM card?

Please update these two doubts,i shall be great full to u.

[Alex McDonald]

Basavaraju

The current PAM card is DRAM based. Yes, there will be further PAM and other flash based developments; the document linked discusses them. Watch this space!

Thanks for commenting.

[Jeff Paul Internet Business]

Great post! My firm does internet marketing for companies and so a lot of my e-newsletter, blog and seminar content revolve around for promotion. Thanks!!!