Now, I have been a fan of using 2.5″ drives in servers and storage systems for a while. But sometimes the rationale for actually pushing such solutions escapes me. The primary reason for this is the cost difference between a 2.5″ drive vs an equivalent 3.5″ drive.
In this article published by Byte and Switch, one of the reasons for moving forward is, “The 3.5-Tbyte SP1224s is Xyratex’s first offering to use 2.5-inch drives, which the vendor is aiming at applications requiring high input/output operations per second (IOPS). By packing 24 2.5-inch drives into a 2U form factor, Xyratex claims to offer double the performance of its predecessor, the 4U, 24-Tbyte SP1424s, which relies on a dozen 3.5-inch drives.”
But this is comparing apples to oranges. Think about it for a second. How can this new solution provide TWICE the IOPS of its predecessor? Both offer support for 24 drives. The predecessor supports 15,000RPM drives whereas the new solution supports up to 10,000RPM drives. Hmm… I’m not quite getting the math. 24 drives is more like two dozen drives– not a dozen. I suppose someone should have proof read the article.
The article goes on to discuss the “green” aspects of the new storage system. But even this is a bit misleading. For example, the SP1424 supports SATA and SAS drives. BTW, there’s no technological reason the new system can’t support both drive types. It’s got to be a positioning thing. Anyway, consider a 2.5″ Savio drive’s capacity of 146GB vs a Barracuda’s 1000GB’s. On a per drive basis, that’s a power consumption difference of 5.5W. Across the entire enclosure, this equates to 88W. However, when you factor the capacity difference, it will require almost 7 enclosures to equal the storage capacity. That 88W advantage becomes rather meaningless when you look at the additional 728W of power that’s necessary to equal the same amount of capacity. That works out to nearly $640/year. So, the positioning is a little bit off.
In all fairness, another aspect of the positioning is towards transactions as opposed to capacity environments. And there’s not much I can do to disagree with that. But let’s avoid the green discussions. If one wishes to have both transactional benefits along with better power efficiency, then SSDs are a better fit than 2.5″ drives. Now, the costs will go up tremendously; but isn’t the environment truly worth it if you are that “green” sensitive?
If you detect a bit of sarcasm, you wouldn’t be wrong. Like I said in the beginning, I’m a fan of 2.5″ drives, but I wrestle with the positioning of them. Ultimately, I believe economics are still a better rule of thumb. And it’s far easier to build fast, low-cost storage solutions with 3.5″ SATA drives than it is with any speed 2.5″ drive. And when you consider that 3.5″ drives can store 4 times the capacity of a 2.5″ drive, I really don’t think density arguments have much weight.
Thanks for the neat solution! But I still believe you’ve missed the mark. However, if you were to come up with a cheap, dense and fast solution, I might be a bit more interested. And I think the positioning would be a bit clearer.
Technorati Tags: xyratex, storage, 2.5″ drives, savio
03/04/2008 at 10:04 am Permalink
Mike, I didn’t get the math either. You have to consider the source of the article – Byte & Switch. If you look at the actual Xyratex press release (http://www.xyratex.com/Company/News/Detail.aspx?ID=225) you will see the correct positioning: Double the performance _DENSITY_ at half the watts per drive. With 24 drives of each (3.5″ or 2.5″) one can expect the same number of IOPS. However, I think you will agree it takes half the rack height with the 2.5” form factor.
On the green front one needs to compare apples to apples. One does not use 1TB SATA drives for enterprise-class, high-performance, high-IOPS applications. If you compare wattage for 15K 3.5″ SAS drives to the 15K 2.5″ SFF SAS drives — you will be comparing 19-21 watts per 3.5″ drive to ~8.5 watts for each SFF drive. This is based on Seagate datasheets for the drives.
Are you suggesting that data centers with high performance needs should not pay attention to reduction of power consumption and better utilization of floor-tile space?
While you are correct that 2.5″ SFF SATA drives are available today, they are 1.5 Gb “laptop” class drives. What Seagate terms “business-class” SATA 2.5″ drives (3 Gb SATA) are not available until late 2008. And you also know that SATA drives are not dual ported, so in an enterprise-class environment additional interposer (MUX) circuitry is necessary to allow dual-host access to a single SATA drive.
03/04/2008 at 10:31 am Permalink
Hey Mark,
Talk about a fast response. You made my day. And you’re right. You have to consider the source of the article. The good news is I at least looked at the statement to ensure it hadn’t come from you
On the IOPS side, I actually somewhat agree with you. Take a look at the following: http://sullego.com/Solutions/text/products/images/netbench-021108.jpg.
We’re clobbering folks with 7200RPM SATA drives, while also obtaining high densities and low costs.
My point is that the best methodologies are not to use smaller hardware. In the end, you end up consuming more power as you have to use more boxes. Bear in mind, data is increasing dramatically each year. We are consuming more storage to house it.
If transactional performance is truly an issue, then there are probably better ways to address the problem. BTW, I’m not really advocating use of NAND drives either. Just advocating smarter system design.
Let’s take for example virtualization. Sure we can connect a Windows server to a Netapp filer via GbE. Or perhaps we can connect via 10GbE. But we could also use a paravirtualized block interface and actually get better performance by eliminating the TCP overhead or even the overhead of Ethernet by directly tying into the system bus.
So, let’s drive performance up. But let’s also drive the power down in a meaningful way. Storage density is best measured by Gbytes per rackspace, not drives per rackspace. And I believe we can in fact have our cake and eat it too.
03/04/2008 at 12:56 pm Permalink
Mike, Impressive performance on the ES320. Scales nicely over the number of clients as well!
I am seeing a lot of innovative methods employed to increase performance of storage infrastructures – including SATA drives. No doubt that within a tiered storage implementation, one will see a wide variety of drive technologies implemented. From SSD to 2.5” to 3.5” (FC, SAS, SATA) to grid architectures to various striping across enclosures to get better aggregate bandwidth.
What I’m suggesting is for IT managers to consider 2.5” for implementations where latency and response time is an issue. This may represent from 1% to 10% of a typical enterprise storage pool. Applications that have small-block, high-IOPS requirements.
If the application is using 3.5” SAS drives today, the rack space and watts consumed can be cut in half by implementing the same number of 2.5” drives.
03/04/2008 at 3:27 pm Permalink
Gotcha! I think we both want the same thing. Unfortunately, it is typically IS group or the system provider who may ultimately determine the architect.
Consider some ILM solutions where 15k SAS drives are used in a mixed array with SATA drives. The premise is the faster drives can be used for faster-access data and then once it’s not as hot, it can be pushed to the slower drives. It sounded good until I did the math. Four SAS drives at roughly 110MB/s for writes is still a lot less performance than sixteen SATA drives at roughly 80MB/s.
The issue for squeaking out performance isn’t in the drive interface, or even in the RPMs of the disks. It’s in the RAID and caching.
Now, latency is inversely proportional to throughput. So as we can drive throughput up, we can drive latency down. This is also a big factor when dealing with lots of concurrent HD video streams.
When I looked at various solutions out in the market, I noticed the phenomena of I/Os falling off. Typically this was in a parity RAID configuration. But not always. For example, I did an analysis using sixteen SATA drives in a RAID0 configuration, but via an LSI 1078 controller. I got a whopping 91MB/s. Now that’s only 5.7MB/s per drive. Most of the time, I end up seeing around 17 to 20MB/s per drive in a RAID5 configuration. This is a good indicator of storage inefficiencies. Where’d the other 60MB/s per drive go?
This is all why I’m advocate of doing things smarter, and in improving efficiencies. Then it doesn’t matter what we use in our systems. If we can make the slower components faster, think about how much faster the more expensive components could be.
Then we can begin to look at power savings. Right now, the inefficiencies are so gross, talking 2.5″ power consumption seems to be a waste of time. It’s only masking the true problem.
04/04/2008 at 7:27 am Permalink
Byte & Switch has made some corrections to their article: http://www.byteandswitch.com/document.asp?doc_id=150035&WT.svl=news1_4