IBM Flash Memory Breaks 1 Million IOPS Barrier

alphadogg writes to tell us that IBM is claiming a victory on the flash storage front. Their new research project "Quicksilver" is claiming data transfer speeds of more than 1 million input/output operations per second (IOPS). "IBM said Quicksilver is two and a half times faster than its own SAN Volume Controller coupled with IBM's DS4700 storage. It would also be two and a half times faster than technology from Texas Memory Systems, which says it has the world's fastest storage with an IOPS rate of 400,000. "

Comments

IBM Flash Memory Breaks?

[Yvan256]

Then I'm not buying IBM flash memory, end of story.

Re:IBM Flash Memory Breaks?

[maxume]

Will Farrell called, he wants you to join his club.

Bit error rate?

[Toffins]

That's very fast. I wonder how low the bit error rate is.

Re:Bit error rate?

[pixelpusher220]

More importantly, how durable is the memory?

Doesn't flash memory have a limit on the # of accesses before it starts to fail?

Re:Bit error rate?

[Surt]

It has a limit to the number of writes, the number of reads is pretty unlimited. The expected average lifetime is similar to a hard drive, and in some cases better.

Re:Bit error rate?

[Surt]

Obviously, there's not a whole lot of independent research out there, but here are some of the claims:

http://talkback.zdnet.com/5208-11408-0.html?forumID=1&threadID=39451&messageID=725468&start=0

http://www.linuxdevices.com/news/NS7676844023.html

http://www.computerworld.com/action/article.do?command=viewArticleBasic&articleId=9007518

Time to market?

[Naughty Bob]

They don't even commit to a date when this might be viable.

Given that current systems are 3 or more orders of magnitude slower than the stated amount, I'm pretty safe in saying that this announcement is meaningless outside of the lab. Kudos, but.... next!

Re:Time to market?

[spun]

If it's in the lab, "Next!" But if it's in the market, "Slashvertisement!" Good old Slashdot, where someone is always ready to shit in your cornflakes.

Re:Time to market?

[Yvan256]

Oh crap, I thought these were Raisin Bran.

I'm waiting for price/performance and benchmarks

[afidel]

While managing to achieve 1M IOPS is somewhat impressive, it's not hard to do for an optimal theoretical situation. Xiotech was showing 500,000+ IOPS from three of their new Emprise 5000 storage shelfs at Storage Networking World this spring, but it was all video and synthetic sequential reads. That same system would only pull about 20K IOPS on the SPC-1 real world benchmark.

Re:It's not a fucking barrier

[Aranykai]

Italics look like this. /This/ /just/ /makes/ /you/ /look/ /like/ /an/ /ass/.

Re:Big deal

[ppc_digger]

Which means it transfers 640 GB per second.

Re:It's not a fucking barrier

[Nimey]

Usenet-isms.

*Bold* /Italic/ _Underline_

But I wouldn't expect you to understand that.

Re:Big deal

[Anonymous Coward]

It only transfers 640KB per I/O operation, tops.

Well, 640kb should be enough for anybody.

soooooo...

[Taibhsear]

How does this translate into normal transfer speed units like MB/s? Otherwise I have no point of reference to tell if I am impressed or indifferent.

Re:soooooo...

[Hecatomb00]

How does this translate into normal transfer speed units like MB/s? Otherwise I have no point of reference to tell if I am impressed or indifferent.

All I know is it is fast. This is a huge win in my book. I am really tired of finishing before my standard hard drive can seek out my porn.

Re:soooooo...

[TheRaven64]

IOPS are a normal measure for server hardware. It's the number of I/O operations per second the device can perform. Most server workloads, particularly database servers, require a lot of small I/O operations per second. With a single mechanical disk, it's pretty easy to work out the number. Pretend the time taken for each read is zero (it isn't, but it's really tiny so we can ignore it). Then take the reciprocal of the seek time. For a cheap disk, this is around 9ms. Google says the reciprocal of 9ms is 111Hz, so a 9ms seek time translates to 111 IOPS. Fast drives have average seek times down at around 4ms, which gives 250 IOPS. So, to put this in perspective, it's as many independent operations per second as 4,000 individual high-end disks, or almost 10,000 cheap ones.

By the way, for a workload with a lot of independent reads or writes you'd be surprised how slow a hard disk is. With a 512byte block (common on hard drives) you get a maximum throughput of around 50KB/s for a random access pattern on a cheap drive, going up to around 125KB/s on an expensive one. Even very cheap flash can do better than this, so for moderate sized databases (a few GBs) with a very heavy access load flash works out a lot cheaper.

Oh, and for reference each of the ops in this test was up to 640KB, giving a maximum of around 640GB/s data transfer.

Re:soooooo...

[TopSpin]

How does this translate into normal transfer speed units like MB/s? Otherwise I have no point of reference to tell if I am impressed or indifferent.

I'll try to help.

MB/s is a measure of IO throughput. Often this isn't the most relevant figure for 'enterprise' storage. Certain applications do a lot of random access IO so IOPS becomes more important than throughput.

Today a typical desktop disk is capable of about 100-150 IOPS. That's a rule of thumb range that varies based on operation size, cache, etc. It works pretty well usually. You can aggregate disks and get almost linear scaling; 12 disks, for instance in a device like this, will give you a maximum of 1200 IOPs, roughly. A common USB Flash device can break 1000 IOPS with certain access patterns.

The second graph on this page illustrates the extreme IOPS advantage of Flash for certain applications. Disks are limited by head actuation and rotation latency. This is why enterprise storage vendors have been pursuing Flash aggressively. That's what this story is all about.

The dream is to host the same IOPS in with an order of magnitude less physical space, power, heat, etc. If you don't need thousands of IOPS (and most PC users don't) then it isn't very interesting. If you happen to run an OLTP system with thousands of reads/write per second it means a great deal.

Re:It's not a fucking barrier

[Just Some Guy]

If it was a barrier, you wouldn't be /able/ to break it.

Yeah.

Re:I'm waiting for price/performance and benchmark

[afidel]

I stand corrected, from the talkback link I followed a trail to an IBM blog with a LOT more details here, and this is the 70/30 SPC-1 benchmark numbers with cache disabled. This is freaking phenomenal performance! The storage is only 4TB, but if you put your logs, flashback, and temp tables on this beast and pinned your busiest tables in ram you would have a screaming OLTP database. I guess it's now just a matter of price, but a rack of x-series boxes with flash card's shouldn't be THAT expensive. Unless IBM asks for a crazy markup it should be affordable for most enterprises (ok, pretty much a given with IBM but still).