Last week, in Computer Technology Review's
article [Tiering: Scale Up? Scale Out? Do Both
], Mark Ferelli interviews fellow blogger Hu Yoshida, CTO of Hitachi Data Systems (HDS). Here's an excerpt:
"MF/CTR: A global cache should be required to implement that common pool that you’re talking about going across all tiers.
Hu/HDS: Right. So that is needed to get to all the resources. Now with our system, we can also attach external storage behind it for capacity so that as the storage ages out or becomes less active we can move it to the external storage. They would certainly have less performance capability, but you don’t need it for the stale data that we’re aging down. Right now we’re the only vendor that can provide this type of tiering.
If you look at other people who do virtualization like IBM’s SVC, the SVC has no storage within it because it’s sitting so if you attach any storage behind it, there is some performance degradation because you have this appliance sitting in front. That appliance is also very limited in cache and very limited in the number of storage boards on it. It cannot really provide you additional performance than what is attached behind it. And in fact, it will always degrade what is attached behind it because it’s not storage, where as our USP is storage and it has a global cache and it has thousands of port connections, load balancing and all that. So our front end can enhance existing storage that sits behind it."
This is not the first time I have had to correct Hu and others of misperceptions of IBM's SAN Volume Controller (SVC). This month marks my four year "blogoversary", and I seem to spend a large portion of my blogging time setting the record straight. Here are just a few of my favorite posts setting the record straight on SVC back in 2007:
Since day 1, SAN Volume Controllers has focused primarily on external storage. Initially, the early models had just battery-protected DRAM cache memory, but the most recent model of the SVC, the 2145-CF8, adds support for internal SLC NAND flash solid state drives. To fully appreciate how SVC can help improve the performance of the disks that are managed, I need to use some visual aids.
In this first chart, we look at a 70/30/50 workload. This indicates that 70 percent of the IOPS are reads, 30 percent writes, and 50 percent can be satisfied as cache hits directly from the SVC. For the reads, this means that 50 percent are read-hits satisfied from SVC DRAM cache, and 50 percent are read-miss that have to get the data from the managed disk, either from the managed disk's own cache, or from the actual spinning drives inside that managed disk array.
For writes, all writes are cache-hits, but some of them will be destaged to the managed disk. Typically, we find that a third of writes are over-written before this happens, so only two-thirds are written down to managed disk.
In this example, the SVC reduced the burden of the managed disk from 100,000 IOPS down to 55,000, which is 35,000 reads and 20,000 writes. Some have argued against putting one level of cache (SVC) in front of another level of cache (managed disk arrays). However, CPU processor designers have long recognized the value of hierarchical cache with L1, L2, L3 and sometimes even L4 caches. The cache-hits on SVC are faster than most disk system's cache-hits.
This is a Ponder curve, mapping millisecond response (MSR) times for different levels of I/O per second, named after the IBM scientist John Ponder that created them. Most disk array vendors will publish similar curves for each of their products. In this case, we see that 100,000 IOPS would cause a 25 millisecond response (MSR) time, but when the load is reduced to 55,000 IOPS, the average response time drops to only 7 msec.
To be fair, the SVC does introduce 0.06 msec of additional latency on read-misses, so let's call this 7.06 msec. This tiny amount of latency could be what Hu Yoshida was referring to when he said there was "some performance degradation". There are other storage virtualization products in the market that do not provide caching to boost performance, but rather just map incoming requests to outgoing requests, and these can indeed slow down every I/O they process. Perhaps Hu was thinking of those instead of IBM's SVC when he made his comments.
Of course, not all workloads are 70/30/50, and not every disk array is driven to its maximum capability, so your mileage may vary. As we slide down the left of the curve where things are flatter, the improvement in performance lowers.
|IOPS before SVC||IOPS after SVC||MSR before SVC||MSR after SVC||Percentage Improvement|
|100,000 IOPS||55,000||25 msec||7.06 msec||72%|
|80,000 IOPS||44,000||12 msec||6.06 msec||50%|
|60,000 IOPS||33,000||7 msec||4.86 msec||31%|
|40,000 IOPS||22,000||5.4 msec||4.16 msec||23%|
|20,000 IOPS||11,000||4 msec||3.86 msec||4%|
Hitachi's offerings, including the HDS USP-V, USP-VM and their recently announced Virtual Storage Platform (VSP) sold also by HP under the name P9500, have similar architecture to the SVC and can offer similar benefits, but oddly the Hitachi engineers have decided to treat externally attached storage as second-class citizens instead. Hu mentions data that "ages out or becomes less active we can move it to the external storage." IBM has chosen not to impose this "caste" system onto its design of the SAN Volume Controller.
The SVC has been around since 2003, before the USP-V came to market, and has sold over 20,000 SVC nodes over the past seven years. The SVC can indeed improve performance of managed disk systems, in some cases by a substantial amount. The 0.06 msec latency on read-miss requests represents less than 1 percent of total performance in production workloads. SVC nearly always improves performance, and in the worst case, provides same performance but with added functionality and flexibility. For the most part, the performance boost comes as a delightful surprise to most people who start using the SVC.
To learn more about IBM's upcoming products and how IBM will lead in storage this decade, register for next week's webcast "Taming the Information Explosion with IBM Storage" featuring Dan Galvan, IBM Vice President, and Steve Duplessie, Senior Analyst and Founder of Enterprise Storage Group (ESG).
technorati tags: IBM, SVC, HDS, USP, USP-V, USP-VM, VSP, CTO, Hu Yoshida, Ponder