Comments on: IBM tests 4-terabyte solid-state drive tech

Big Blue announces solid-state drive technology that achieves high speed and power savings, even if only in the labs so far.

[whizkid454]

And the price is somewhere in the 5 figures....

[timber2005]

It's not in the consumer level though. Its an "array" of multiple drives that totals to 4TB.
Either way, I'd say somewhere in the upper 5th if not 6th figure area.

[OSSG]

LOL- 5 figures? Would be nice... with the price of flash drives, try 7 figures. Then add in the SVC controllers and licensing.

[James6501]

"SSDs are being supplied by Fusion-io"

How would Fusion's ioDrives fit in IBM's storage? I thought they were SSDs made for servers?

[RompStar_420]

just when the first hard drives came out, they were super expensive, with time this will be reduced to pennies on the dollar

[tech_crazy]

So, all IBM is doing is testing them? And this is headline news?

[aaroberts]

EMC has already been using SSD's since the beginning of the year. STEC makes them.

[orbist]

We are indeed using FusionIO cards to build a storage controller units that presents storage onto a fibre-channel network, in this case storage is presented behind a large SVC cluster.

The workload benchmarked was 70/30 4K all miss which is somewhat representative of a typical database workload.

If people have actually played with SSD devices, they will know that the claimed X0,000 reads and Y,000 writes are all very well when doing 100% read or write at say 512 byte transfers, but bump up the transfer size, and throw in a mixed workload and you will prune the men from the boys... i.e. IOPs will be no where near those claimed by manufactures...

While EMC has been shipping SSDs since the 2Q, the question I'd ask is how many DMX quadrants would be needed to provide this same level of IOPs... many many... which defeats the power and spaces savings that SSDs can provide.

Anyone interested in more details of what we have been doing see my recent blog post :

http://www.ibm.com/developerworks/blogs/page/storagevirtualization?entry=1m_iops_from_flash_actions

[hastingd]

120 drives in Quicksilver. 120 drives in the skunkworks first generation XIV Storage System. Surely a cooincidence...

[James6501]

Assuming these are truly off the shelf Fusion SSD's like IBM claims. How many Fusion cards does it take to get this level of performance? I assume there are a whole lot of them and if there are a bunch, not to sound defensive, claiming such a low latency number with mind boggling IOPS to boot sounds a little to good to be true. Not even EMC can get that kind of IOPS performance with such a low latency number. Assuming this all stands up to scrutiny, what does Fusion & IBM know that no one else seem to know when it comes to SSD storage?

[orbist]

Scale-out controller. Its as simple as that, as I alluded to in my blog above, the problem with just sticking form factor drives into an existing array controller is that the controller itself will limit the performance. Its not been designed around the performance and latency characteristsics of flash, exactly the opposite. All of todays big iron controllers have been designed to make a sequential device, (HDD) look like a random access device (from a latency point of view) The head has to seek to the right LBA and then do the work - with flash its all at silicon speeds, and no seek latency is added.

For write workloads, storage controllers have added cache and the like to hide such latency, but that becomes less improtant with flash. Also consider a (large) FC-AL loop of HDD, say 128 HDD, then the dual loop config needs to only support say 128x350 IOPs ~=45K iops (assuming all HDD are flat out - which is unlikely to be possible) so thats one flash device... There are likely to be bottlenecks further up the system too... So while it does help improve latency by using existing controller devices, it cannot max out the devices, not can it provide the best latency.

With a modular system like IBM SVC, then you have 50us of latency top to bottom of the stack, and with the flash devices typically giving 400us under heavy load, you can keep to around 700us for the entire system, host -> SAN -> SVC -> flash.

Monolithic controllers are fine until they reach their limit, then you need to bring in another monolith and scale out in islands, with a modular scale out box, you simply add more processing nodes (SVC) to the cluster, and more storage behind the cluster and still have a single image and control point.

For this config we used around 40 FusionIO cards.

[wobrien2]

It is great to see more real world I/O being done in this test by using a 70/30 read/write mix with 4K I/Os. However, enterprise storage solutions also use data protection features like RAID.

Was this testing also done with some form of data protection?

[thestorageanarchist]

For a balancing perspective of this science experiment, check out my blog: http://thestorageanarchist.typepad.com/weblog/2008/09/1023-its-just-a.html

[bitbucketio]

orbist : I assume you're using the IOdrive without the IOsan port. How many IOdrives were in your test?
Are there any plans to use the IOsan in your storage solutions independent of the SVC?

http://www.fusionio.com/PressDetails.aspx?id=44