You know when you hear your own company say something over, and over again – sometimes one can have their own doubts.
I remember the first time Donatelli (formerly at EMC, now at HP) told me: “in 2-3 years, there will be no place for a high-performance rotating rust disk, we’ll only have solid state and very large slow magnetic media” (that was about 1 year ago).
I was a bit skeptical.
I was also talking to a customer about it yesterday, and we debated about cross-over points (not whether it would happen, but just timelines).
This just clears away any of my skepticism.
So – this article is on Anandtech – one of my favorite non-storage related IT sites. If you don’t want to read the whole article, jump to this page.
Or – let me summarize it for you. Read on if interested, or if after reading the article the “tada!” epiphany doesn’t jump right out at you.
Here’s the standout thing in the article:
- Kingston 40GB drive (the purple line on the chart from Anandtech), based on the 35nm Intel X-25M
- Price: $115 before rebate ($85 after mail-in rebate, but am not going to assume that)
- delivered 4000 4K IOps random writes.
- delivered 7500 4K IOps random reads
Don’t get the big deal? First – look at the odd ones out. Those are fast spinning magnetic media disks.
Still don’t get it? Ok, let me make a comparison. Sure, a 1.5TB 7200 SATA drive can be bought for $115. But it will do 80 4K random write IOps.
- Expressed as GB/$ = the SATA drive is a 38x better deal
- Expressed as IOps/$ = the SSD drive is a 50x better deal (for write IO workloads, for reads, it’s 93x better). The SSD delivers 34 IOps per dollar. The SATA drive deliver 0.69 IOps per dollar.
To match the random read IO performance of that $115 Solid State disk, you would need 50 of the 1.5TB SATA disks.
But surely, if you were looking for performance, you wouldn’t use the SATA disk, right? You would probably use a 15K RPM FC disk. Those cost about $1000. They do about 200 random write IOPs. So, you would need 20 of them to do what that $115 SSD could do. That’s 0.2 IOps per dollar – or 170x more expensive than the SSD on a IOps/$ basis. Oh, you think SAS 15K drives are a better deal? They are – than FC disks. A 15K SAS disk on Pricewatch costs about $210, and they also do about 200 IOps. that’s 0.95 IOps per dollar – or 37x more expense than the SSD on a IOps/$ basis.
Also – all the “more for less” items are purely expressing acquisition cost. Solid state is also orders of magnitude better in power efficiency and density.
Now, sometimes, performance is focused on bandwidth (MBps) – this is usually for sequential IO workloads – very common in backup to disk use cases, and in those cases, spinning rust does OK, and often faster pipes (10GbE for example) are “efficiency technologies” – enabling you to get more for less, and so is big SATA and fantastic dedupe (we do them all of course :-)
For most production workloads in the 4K to 64K average IO size, performance it tends to be gated by IOps.
Ok – fast forward this just a LITTLE bit. a 32GB commerical MLC-based SSD cost $900 near the beginning of this year. Now, it’s 10x cheaper.
Today, there’s only one manufacturer that has a LOCK on the “enterprise” SSD (or Enterprise Flash Disk) market – a company called STEC, and we can put them in all our arrays. We’ve sold out for the last 6 quarters. I also got some great customer feedback after I first observed that (this is a customer who use EMC arrays with solid state right now). From DaveFW on twitter:
“RT: @sakacc re:SSD Its had a HUGE impact on our production oracle database already”
“RT: @sakacc re:SSD we had jobs that ran 2 days take 10 hours now with no impact to users”
But today, without automated tiering, and with the $/GB being skewed against SSDs – they are reserved for applications that have low capacity, but high IOps workloads. Theses aren’t unusually (they are everywhere), but they aren’t universal.
There is more to the story, however. EMC is one of Intel’s biggest (in fact I believe THE biggest) non-server OEMs. We use Intel’s CPUs everwhere, and are now nicely riding the Intel Nehalem roadmap. Pat Gelsinger is here now which also helps. I can tell you for a fact that the Intel Flash folks are consistently making progress towards enterprise-class SLCs. Samsung is pushing hard too. BTW – I’m assuming no “star trek” technology like phase change memory, this will happen with the basic technology that exists today.
What will it mean when a SSD has the same $/GB and 100x better $/IOps than a 15K 500GB SAS disk?
I no longer doubt what Donatelli said.
I also don’t doubt that the arrays of the future will all need the ability to leverage a combination of very large SATA and very large solid state storage. The curve on this is a classic innovators dilemma example of a disruptive technology coming into an existing market. The external forces will simply be too strong to ignore.
Exciting times in storage land – exciting times!
I’ve got to get myself one of those kingston drives!!!!
Would be fun to get input from people on when they think the inflection point (volume of solid state used in enterprise servers and enterprise storage has a higher volume than high performance spinning media) will occur… Please, comment!