Fusion-io: Can you say Super Speedy Database?
Fusion-io is bringing a whole new meaning to SSD – Super Speedy Database.
Last week I had lunch with Sumeet Bansal, Fusion-io’s Principal Solutions Architect. Fusion-io is a producer of enterprise-class SSDs (Solid State Drives, as opposed to HDDs – Hard Disk Drives). Fusion-io applies flash memory to large-scale enterprise storage products like Storage Area Networks. I first met Sumeet when he was VP of IT at Wine.com.
Part of our discussion was about their products, and how they differ from the competition. SSDs can be very fast. They can also be pretty rugged, as they have no moving parts. They are more expensive per GB than conventional hard drives – at least for now. SSDs eventually wear out – each spot on the NAND Flash Chip can be exercised a finite number of times before it stops working.
Key points regarding Fusion-io’s ioDrive:
Performance of NAND Flash Chips is extremely fast. Read latency is an order of magnitude faster than conventional HDDs. But the Fusion-io ioDrive removes the IO bottleneck and makes the cpu utilization more efficient. Fusion-io decided to adopt a PCIe architecture rather than dealing with the multiple layers of protocols necessary to implement a SATA interface. I reviewed performance results at HotHardWare, and it’s clear that they are getting a big performance advantage.
They have the wear under control. Their controller spreads the “wear” across the entire drive, giving them excellent life (1.2 million hours MTBF!)
The rest of our discussion was around how the ioDrive could be used most effectively on different platforms we work with a lot at DesignMind. There is work to be done regarding Best Practices for implementing SSD technology for SQL Server, SharePoint, Exchange, and other key Microsoft Server tools. We look forward to working with Fusion-io to help them develop Best Practices, Performance Benchmarks, and Case Studies.
Sphere: Related Content6 comments
6 Comments so far
Leave a reply



Hi Mark -
This is very interesting. How about a presentation on this topic for BAADD’s Silicon Valley SQL Server User Group? Stuff like: more detailed mtbf comparisons, failure mode comparisons, any application eccentricities, cost comparisons. I assume you’d: put logs on it, and put tempdb data on it. And I assume you’d RAID1/10 it too, right ? Any other trick application approaches .. ?
Steve
Thanks for your ideas, Steve. We definitely plan to have Fusion-io come in to speak about SSDs, benchmarks, best practices, and SQL Server performance.
There is more info on SSDs at Steve Jones’ blog:
http://www.sqlservercentral.com/blogs/steve_jones/archive/2009/07/14/the-fusion-io-ssd-review.aspx
Also, Wesley Brown promises us some performance results soon.
http://www.sqlservercentral.com/blogs/sqlmanofmystery/archive/2009/05/29/fusion-io-s-on-their-way.aspx
I’ve been a proponent of SSD and normalized databases for years (Texas Memory Systems has been building DRAM SSD for years, and now has flash parts). It’s only since STEC got attention as an EMC supplier that the enterprise world has taken notice; the Intel X-25 parts for the rest of us. It took STEC nearly 4 years to get shipments that were noticeable. Whether Fusion-io can play in the same yard is another question, since they haven’t gone FC; although I have a memory that HP has an FC rack adapted to Fusion-io parts.
The key to getting the most out of SSD is normalized data. They’ll always be way too expensive for a one for one swap for HDD. EMC is going the tiered storage route with SSD facing the application, backed by HDD and EMC firmware/software. I wish them well, but the advantage of SSD lies in data shrinking with high normal form data redundancy removal into pure SSD. We shall see.
Running a large online gambling facility – have 2 dozen or so on order… rather looking forward to it
Tested 3 in a stripe, and they are insanely fast – 3GB/s read speeds…
Had processes that took several minutes complete in seconds, sample batches of 1000 ‘financial transactions’ stress test completing in under 100ms instead of the 500ms on DAS or 700ms on SAN.
Reorganising data – stripping out archivable data from a 1.5TB database and shrinking the files to 450GB, while moving all data to new filegroups… estimated (after tests) at well over 30 hours on a 55disk R10 SAN – down to under 3 hours.
Can’t say how much I am looking forward to getting this in production
you can check the IO performance improvements with Firebird database on FusionIO
http://www.firebirdnews.org/?p=3381
For my realtime needs I would switch to ssd for the critical random access io (like heavy self joins)
[...] Fusion-io is a producer of enterprise-class SSDs. Their silicon-based storage architecture known as ioMemory applies flash memory to large-scale enterprise storage products like Storage Area Networks. [...]