thinking sysadmin » ssd

Links 7/22/2008: NetApp and Flash

Andy — Tue, 22 Jul 2008 19:42:10 +0000

Flash Forward – Jay Kidd, CTO of NetApp blogs that “NetApp is in the process of certifying enterprise-grade SSDs that you can use in our existing storage shelves.” No dates or pricing announced yet, of course, but he does make an excellent point about SSDs in storage arrays: “For the next few years, you won’t be using a lot of flash capacity in your systems, not just because of the costs. At 10x or more the IOP rate of hard disks, it only takes a small number of SSDs in disk slots to saturate the performance of the array controller. It’s like trying to fly a model airplane in your living room – you’ll run into a system performance wall long before you hit capacity limits. This is another reason that flash as cache is economically efficient – it puts the necessarily small amount of very fast storage at a point in the architecture where you can best exploit the performance.” Not unlike how Sun suggests using SSDs with ZFS. (Seen at Blocks and Files.)

Catch-up Links, 7/9/2008

Andy — Thu, 10 Jul 2008 18:47:03 +0000

There’s nothing like a long summer weekend followed by an on-site consultant to keep you from updating your blog. But on the bright side, I didn’t have to link to the notebook SSDs are dead – no they’re not kerfluffle.

NetApp finds NAS could mean ‘never accessed storage’ – “According to a USENIX presentation, 90% of data on NetApp’s networked storage systems was untouched over a 3-month period, raising the issue of whether it would be better placed on cheaper storage.” I would find some irony in that cheaper storage being tape.
zfs un-benchmarking – “Our rationale for testing was to finally get some numbers that we can provide to users/customers about real zfs performance. There is a huge amount of (largely uncontested) information (emanating mainly from Sun and its agents) that zfs is a very fast file system. We want to test this, on real, live hardware, and report. Well, we can’t do the latter due to Sun’s licensing, but we did do the former. Paraphrasing Mark Twain: ‘Rumors of zfs’s performance have been greatly exaggerated.’” When Joe Landman blogs about performance, I take what he has to say seriously, but given the stability problems he notes, I wonder if – as he suggests – that driver issues are a factor here, and we’re not seeing a generic ZFS issue. (Seen at InsideHPC.)
Self-protecting archive for SharePoint – “A new archiving product from BridgeHead Software automatically moves older, infrequently-accessed SharePoint items to cheaper archive media and cuts down the SharePoint backup burden.” HSM for SharePoint, apparently? Sounds interesting.
Since NFSv4 is Stateful It Must Be Less Robust, Right? – “The short answer is no.” Interesting summary of how locking works under NFSv4; although I haven’t used NFSv4, this sounds like a massive improvement over previous versions – can I get the time that I spent debugging locking problems on Linux NFS servers back now?
Storage Opens Up – Sun releases new JBODs and upgrades Thumper’s hardware. Given that I heard rumors of the Thumper expansion shelves (J4500) maybe a year ago, I’m surprised it took them so long to come out. And is it just me, or do the J4200 and J4400 look a little like someone else’s boxes rebranded?

What I Think You Ought to Read, 6/20/2008

Andy — Fri, 20 Jun 2008 19:14:52 +0000

Red Hat adopts KVM: what happens to Xen now? – I work in a VMware ESX shop right now (other than all those Solaris Zones and FreeBSD Jails and OpenVZ VEs, that is) – no Xen or KVM. However, given the serious pain in the butt that timekeeping is in Linux guests on ESX, I’ve been sorely tempted to look at running Xen for some Linux virtuals under CentOS 5, in the hopes that this isn’t a problem there. Guess I’ll hold off on that now. (Yeah, I’ve read all the docs, and Linux time sync generally sorta kinda usually works until it doesn’t – it just shouldn’t be that much of a flail.)
10 Gigabit Ethernet and VMware – A Match Made in Heaven – Lengthy and interesting article on 10GbE, VMware, consolidation and datacenter Ethernet – dismissive of Infiniband’s chances of becoming the One True Network Fabric.
Blocks and Files: Between a server or storage array place – More commentary on Chuck’s commentary on HP’s flash announcement. Quote: “Another aspect of this is that a flash SSD cache for a servers needs to plug in to the server’s bus and the supplier doesn’t have to worry about getting a Fibre Channel interface onto flash SSDs which is needed to plug them into existing Fibre Channel slots in a storage array. STEC has an effective monopoly on this (with EMC having its own mini-monopoly because of its exclusivity deal with STEC which ends in a few months) until Emulex’ SSD-tweaked SATA-to-FC bridge chip becomes available at the end of the year. ” Which is probably why NetApp recently announced this instead of flash drives.

EMC’s Flash Blind Spot

Andy — Fri, 20 Jun 2008 13:33:02 +0000

Chuck’s got another, uh, thought-provoking blog post up, More Examples Of Why Server Vendors Just Don’t Get Storage, surely intended to ruffle a few feathers. And he does raise some really good points: Most server vendors need more of an SSD strategy than just making a flash drive an option (it’s how you use it, not that you have it!). And as big a fan as I am of ZFS and Sun’s storage options in general, to win in the “enterprise” (and not just, say, HPC) Sun needs to pull everything together into Solaris (from OpenSolaris) and make it less of a DIY operation.

But here’s where I think EMC’s missing something – their flash blind spot, if you will: Price-performance. Sure, I recognize that EMC has chosen to enter the market from where they compete best – the high end. And I realize it’s inevitable that flash will move down-market in EMC’s arrays (really: flash in the CLARiiON line is a “when,” not an “if”). But EMC’s flash options are right now way too expensive for most shops, while, on the other hand, there’s a good probability I’ll have servers with SSDs in them by then end of this year. I’ll just need a strategy on how best to use them with my operating systems and applications to maximize their benefit. Unfortunately for HP and Dell, they’re not offering a strategy that I know of, they’re just offering the drives. Sun, on the other hand, has a good use case: ZIL and L2ARC on flash – now it just needs to get back-ported from OpenSolaris into Solaris.

Web Pages My Boss Should Read, 6/19/2008

Andy — Thu, 19 Jun 2008 21:12:56 +0000

HP servers to get flash memory I/O boost – It looks like SSDs will become a common option on server hardware; the real issue will be how to adapt your operating system/file system/application to use them optimally.
BlueArc’s Titanic benchmark holed – A storage vendor using a non-real-world configuration to optimize benchmark results? Really? Say it ain’t so!

Flash – A tale of three companies: EMC, NetApp and Sun

Andy — Fri, 13 Jun 2008 14:17:56 +0000

There’s been a lot of noise from the storage industry about flash recently – in particular, noise from EMC and Sun, both of whom recently announced storage products using flash, EMC in January and Sun earlier this month. Below are my thoughts on what EMC and Sun are doing, as well as what NetApp might do. Since I see a fair amount of visitors from all three companies here, if I’ve got something about your employer wrong, please correct me in the comments.

EMC’s approach seems to target the very high-end. I don’t have any EMC gear to play with, but if I understand correctly, their flash drives are pretty much plug and play with their DMX4 arrays, matching the physical form factor of their existing disk drives. Once the drives are in the array, you can do anything with them that you do with standard drives – these drives are just a lot faster. Sounds very much like a no-holds-barred-if-you-need-speed-this-is-how-we’ll-give-it-to-you – with a price to match the performance.

I’d love to have some of the EMC flash gear, but I doubt I’ll ever see this iteration. I work with NetApp storage arrays (at least for this budget cycle…), and from what I understand about the pricing of the EMC kit, it’s targeted way over the head of my employer. So will I see something similar from NetApp? Unfortunately, it doesn’t sound like they’ll have anything out in the near future. A couple years ago, Dave Hitz was almost dismissive of flash, answering “I don’t think so” and “I don’t care” to the question of whether flash would replace disk drives – and he has some good points, mainly that the storage and management layer is ultimately more important than the physical media one uses. More recently, Jay Kidd, NetApp’s chief marketing officer was quoted by ZDNet as saying that flash will become “an integral part” of enterprise storage – but there are no flash product announcements from NetApp that I’m aware of.

Given NetApp’s storage system architecture, this seems particularly unfortunate to me. SSDs seem like they could fit very nicely as a cache device between the controller and the conventional disk drives. I imagine that just a couple SSDs could give a substantial performance boost to an array for a very moderate price. Additionally, I don’t see a reason why NetApp couldn’t allow you to make an aggregate out of a shelf of flash drives for servers that need absolute top speed. Will they do either of these? I hope both, but we’ll have to wait and see.

(With their newly-announced Performance Acceleration Modules, NetApp seems to be acknowledging the value of an additional caching layer, although the PAM is for read performance only.)

Sun’s flash announcement is focused on the use of flash drives in their server hardware. Specifically, according to an Intel Developer Forum presentation, at least one use case Sun is targeting is ZFS, where direct-attached flash would be used in a cache role for specific ZFS reads and writes (in the second-level Adaptive Replacement Cache – L2ARC – and ZFS Intent Log – ZIL -pools – see page 11 of the above slides). The presentation suggests a possible configuration, comparing a hypothetical Sun X4450 with seven 10k RPM SAS drives and one with two SSDs and five 4200 RPM (did they mean 7200 RPM?) SATA drives. For almost the same price, the SSD/SATA system delivers substantially better Read IOPS and slightly better write IOPS.

Which is all good and nice – I do love a slide with reassuring bar graphs – but Sun won’t be shipping servers with flash drives until later this year. Additionally, the support for separating the L2ARC and ZIL is only in OpenSolaris, not Solaris – I haven’t seen word on when it will be backported. I also wonder about what effect single ZIL and L2ARC drives, as used in the scenario in the presentation, would have on a machine’s availability; Sun could presumably address this with some “best practices” documentation by the time they ship servers with SSDs. All that aside, I like what I see from Sun so far, particularly that they’ve found a use case for SSDs at a reasonable price point.