Fusion-io had the PCI-e flash card market all to themselves for the longest time – but no more. Texas Memory Systems, a stalwart in the DRAM-based SSD market, has announced a new product, the RAMSan-20, a 450 GB SSD on a full height, full length PCI-e card.
With 450 GB of usable single level cell (SLC) flash onboard, the company is aiming at enterprise-class users who put a premium on reliability and availability. SLC is typically spec’d at 10x the writes of cheaper multi-level cell (MLC) flash. Like Fusion-io, TMS provides a thin block-device driver.
With an ~$18k list price, the RAMSan card isn’t for gamers. But the large capacity offers a lot of go for servers large enough to drive it at 120k random reads. That should appeal to the TMS customer base and reseller channel.
The StorageMojo take
This won’t be the last PCI-e flash card announcement this year – expect another tomorrow. And maybe next month. The obvious rightness of putting flash on PCI-e has several vendors panting to get to market.
Vibrant competition will develop this year. Given the economic tide the ability to make a paid-for server act like a new and bigger server will be very attractive. In 3 years we’ll all be wondering “how did we manage before flash card SSDs?”
Courteous comments welcome, of course. I shot a complimentary video of TMS president Woody Hutsell discussing flash and DRAM SSD last year. I’ve done some work for Fusion-io, including this video 4 months ago. Both worth a look.
Check this out:
http://blogs.sun.com/ahl/entry/ssd_announcement
http://www.theregister.co.uk/2009/03/11/sun_ssd_servers/
Since we’re sharing links, this is an interesting link on SQL using a Fusion card.
http://www.jdconley.com/blog/archive/2009/03/12/iodrive-changing-the-way-you-code.aspx#random
I’ve read Anand’s reviews of some of the SATA style SSDs. He points out that the controllers on those devices play games to keep write throughput up until about 60% of the storage capacity is used. After 60%, the controllers need to re-org info to create pages to write to. This leads to reduced throughput over time, unless you periodically use BIOS level command to reformat the SSD.
Do the non-SATA devices, like the RAMSan-20, do anything smarter than “put off re-org until it can’t be avoided”? Seems like they could do things in the background, although emulating block-level devices doesn’t seem to offer much hope.
RC, there’s quite a bit of work involved to make flash look like disk: write leveling; garbage collection; table maintenance. There are big tradeoffs in bandwidth, speed, buffer and block sizes – as well as cost. Cheap drives don’t have the chops to handle all that gracefully. Learn more here.
I expect that the TMS design is stable under a variety of workloads. But maybe they’ll comment.
Robin
Thanks for the info and the link.
The paper mentioned has moved here: http://research.microsoft.com/pubs/63596/USENIX-08-SSD.pdf