By Calvin Zito
Earlier this week, HP announced new 2.5" small form factor (SFF) 300GB SAS drives. You can read a bit about them in a blog written by my colleague Sonia Mathur on the ProLiant server team. Here's the link to her post: http://www.communities.hp.com/online/blogs/reality-check-server-insights/archive/2008/12/02/small-form-factor-drive-storage-doubles.aspx. We'll leverage the SFF drives in several of our StorageWorks products like the MSA, storage blades, and the All-in-One Storage blade.
You might also be interested in an article written on SearchStorage.com about the drives: http://searchstorage.techtarget.com/news/article/0,289142,sid5_gci1341037,00.html
Have a good weekend!
Last month we announced a world record SPC-2 number by the XP24000. At the same time we extended yet another challenge to EMC to join the rest of the world in publishing benchmarks. They continue to decline the offer, arguing “representativeness”. I thought I’d clear up the “representativeness” question.
EMC’s argument that this XP is too costly starts from the assumption that SPC-2 only represents a video streaming workload. To quote, “128 146GB drive pairs in your 1152 drive box? A pure video streaming workload?” We actually see a widely diverse set of workloads used in the XP. The power of having both SPC-1 and SPC-2 benchmark results is that they provide audited data that applies to almost any workload mix a customer might have. But if one had to pick a most common workload it would probably be database transaction processing by day, then back up and data mining workloads joining the transaction processing by night. SPC-2 models the back up and data mining aspects, with SPC-1 representing the transaction processing. SPC-2 is about a lot more than video streaming.
When people need bullet proof availability and high performance for transaction processing they turn to high end arrays like the XP24000. It’s probably the most common use for a high end array. Our data indicates that on average the number of disks in an initial XP purchase is right around the 265 in our SPC-2 configuration. Some of those won’t have the levels of controllers in the SPC-2 configuration. But an increasing number use thin provisioning. In those cases they will often get all the controllers they’ll need up front, delaying the disk purchases as you’d expect with thin provisioning. So the configuration and workloads look pretty representative.
Then consider a real use of the benchmark. A maximum number is key in assessing an array’s performance. Below that you can adjust disks, controllers, and cache to get fairly linear performance changes. But when you reach an array’s limit, all you can do is add another array. So once you know an array’s maximum number you know its whole range of performance. By maxing controllers we provide that top end number, giving the most useful result. For sequential workloads like back-up and data mining maxing disk count isn’t necessary, whereas it generally is for random workloads like transaction processing.
Now let’s discuss how one might use XP’s SPC-2 results. Let’s say you need a high end array for transaction processing. The most common case we see requires backup and data mining operations at night in a limited time window. Since the XP’s SPC-2 result is twice that of the next closest high end array, you can expect it to get the backup and data mining done with half the resources of the next fastest array. But with SPC-2 you can go further. You can look up the specific results for backup and data mining workloads which are around 10GB/s for the XP24000. Knowing how much data you need to backup and mine you can estimate how much of the system’s resources you’ll need to get those things done in your time window and therefore what’s still left for transaction processing during that window. You can scale that for the size array you need for transaction processing. And you can compare to other arrays that have posted results. All using audited data before you get sales reps involved.
SPC benchmarks are all about empowering the storage customer. XP24000’s SPC-2 result is important to the most common uses for high end arrays, as well as for less common uses like video editing. The configuration we used looks pretty typical, with choices made to make the result most useful to customers. The cost is pretty typical for this kind of need. At HP we expect to continue providing this kind of useful data for customers. And our challenge to EMC to publish a benchmark result still stands, though they’ll probably continue inventing reasons not to.
EMC, we're once again throwing down the gauntlet. Today the XP24000 put up the highest SPC-2 benchmark result in the world. The top spot for such demanding workloads as video streaming goes to the XP. Once again, your DMX is a no show. And once again we challenge you, this time to put up an SPC-2 number. Every other major storage vendor is now posting SPC results. Every other major storage vendor is now starting to give customers standard, open, audited performance results to show what they've got. You remain the only vendor keeping your product performance behind a smoke screen of mysterious numbers and internal testing. We challenge you join us in the world of openness and let customers quit guessing at how low the DMX's performance really is!
By Jieming Zhu
I will continue on with part 2 now. Hope you enjoy and please share your thoughts!
Claim #3: Because SSD does not use any mechanical moving parts, it's inherently more reliable than HDDs.
This is probably the biggest misnomer surrounding around SSD's enterprise adoption. Quite to the contrary, because of the lack of enterprise usage, the reliability of SSD is unknown and confusing to most users, even vendors. While we have decades of experiences of how to measure, protect and improve the reliability in the HDD world, the work on the SSD side has barely started. Often times, we hear conflicting definitions about MTBF when it comes to SSD. Do you mean useable capacity or raw capacity of an SSD drive? Do you cite the underlying flash chip number or you have done drive level testing/validation? Do customers know that MTBF actually has nothing to the write endurance? Speaking of write endurance, there are so many different parameters that flash vendors tend to tweak to make their numbers better. Every time when we hear a write endurance number, do we know the actual formula used? What's the page size? What's the I/O block size? How are they mapped? In addition to wear-leveling, what about the read disturbance? How do we handle the bit error? How do we deal with the data retention problem?
Where it matters most, the reliability of enterprise SSD poses many questions, with so few answers...
Don't take me wrong. I am actually a firm believer that SSD, or rather, the general solid state storage technology, brings tremendous performance/cost benefit as well as the energy and space efficiency to data centers. Further more, I also believe that it is a potentially a disruptive technology that will bring some fundamental changes in how storage subsystems are designed, how servers are designed and how OS's and applications should change for the NVRAM access.
As an industry, we have the responsibility to proactively push the technology (we have endless challenges in that front alone!) and responsibly market it to help our customers adopt. Over hyping can only cause the backlash that ultimately hurts us all. It would be particularly a shame for such a promising new technology as SSD.