The unfolding TMC RamSan vs IBM SVC story has unfolded some more. It now transpires that IBM's SAN Volume Controller used solid state caching in its SPC tests.
IBM's Tony Pearson admits that there is a write cycle limitation with flash but TMS' Woody Hutsell points out that TMS uses DDR RAM, not flash, and so the write cycle problem doesn't exist.
The contributions are below.
Tony Pearson writes: "It's true. The SAN Volume Controller (SVC) uses SDRAM with battery backup. The tested configuration runs an 8-node cluster, each node has dual Xeon processors and 8GB of SDRAM memory. The latest SPC-1 benchmark was 155,519 IOPS for 12TB configuration. By comparison, Texas Memory Systems tested their RAMsan-320 which came in at 112,491 IOPS for a small 68GB configuration.
As for SPC-2, the latest SVC benchmark is 4,544 MB per second. This workload tests large file transfers, large data base queries and Video on Demand workloads. SPC-2 doesn't show any results from Texas Memory Systems, and I suspect they may not have a large enough configuration to run it.
I don't know any details of the inside technology of the RAMsan-320, but it might have fewer processors, slower memory, fewer host adapters, slower ports, don't know.
The SVC has been around since 2003, we are on its fourth generation, has over 2,200 customers, and was engineered for optimal performance.
TMS' Woody Hutsell writes:
A few comments:
1. I think your post, insofar as it refers to Hybrid drives (mix of flash memory with hard drives), is accurate.
2. IBM does indeed have the top SPC-1 results (by the way I am delighted to just see SPC discussed in a blog). I do appreciate your thoughts on why the Texas Memory Systems results are lower than IBMs, but I think you missed the real reason... the type of servers used in the testing. If we could possibly borrow the server that IBM used in its testing, I think we could make that fine IBM server look even better than your also excellent storage makes it look. Can you work that out for us?
3. Not all solid state disks have write performance issues or write wear issues. Systems made with DDR RAM, such as ours, have dramatically better write performance than disk or flash and do not have wear issues like flash memory.
4. Your first comment to max out memory in the server has two possible weaknesses. First, server memory is worthless where write performance is the number one bottleneck. Second, server memory in capacities similar to what solid state disks, such as ours, can provide can be more expensive per capacity, cannot be easily shared among multiple servers simultaneously, and almost always have to be thrown out when the server is replaced. An external solid state disk often has a lower cost per capacity than large memory in servers, can work in a SAN connected to multiple servers simultaneously, and does not have to be replaced if the server is upgraded.
5. Regarding your second point, adding performance tiers for some businesses equals complexity. For other businesses, adding tiers equals performance and cost-effectiveness. The whole promise of ILM is that we don't have to be stuck with just three tiers of performance. In fact, I think your SVC is an excellent way to open up the enterprise to multiple storage tiers and yet manage this complexity from a single appliance.
Thanks again for the discussion.
Texas Memory Systems
Woody Hutsell added
1. The IBM SVC uses SDRAM cache not a flash cache. SDRAM and DDR RAM (the memory in the RamSan) are basically the same (technically DDR RAM is a type of SDRAM).
2. The difference between the SVC implementation and the TMS implementation is that the SVC uses the SDRAM as cache (a way to buffer reads and writes from memory based on a caching algorithm). The RamSan is used as a solid state disk (a storage target with memory as the storage media).
3. Also, it is worth keeping in mind that the SVC is more of a SAN virtualization appliance than a storage device. It sits in front of other storage systems providing volume management and data protection services, etc.