Solid state disk vendor Texas Memory Systems has claimed an SPC benchmark world record using industry standard server and network hardware.
SPC stands for the Storage Performance Council, a vendor-neutral body which has devised several benchmarks to compare and contrast storage system performance in various scenarios. The SPC-1 benchmark is based on a workload designed to be representative of typical online transaction processing (OLTP) environments, such as airline ticket sales. It emphasises storage read and write access over sheer capacity or streaming I/O performance.
Texas Memory Systems (TMS) provides DRAM-based solid state disk arrays that look like hard drive arrays to storage-using applications, but respond many times faster to I/O requests. There is no waiting for a disk drive's head to move to the right disk track when an I/O request is received.
The RamSan-400 can have 32- to 138GB capacity and uses either Infiniband or 4Gbit/s Fibre Channel to link to servers. The test set-up used “white box” servers with 4GB of memory to operate the SPC-1 benchmarking application, QLogic QLE2462 host bus adapters and SANbox 5600 fabric switches.
In the SPC-1 benchmark, it delivered a 291,208.58 SPC-1 IOPS (input/output requests per second), with an average response time of 0.86 milliseconds per I/O request.
In comparison, an IBM DS8300 Turbo hard-drive array-based system recorded 123,033.40 SPC-1 IOPS, less than half the TMS number. TNS, Sun, IBM, Dell, Fujitsu, HP and 3PAR have all submitted systems to the SPC-1 benchmark.
EMC has recently announced a DMX-4 drive array which can use flash memory-based SSDs instead of hard drives. It is expected that these will be significantly faster in responding to I/O requests than a hard drive-only DMX-4. Flash memory SSDs are expected to be 30 percent more expensive per GB than the fastest hard drives. A RAM-based SSD is more expensive, and faster, still and, to that extent, comparing its performance to a hard drive-based array is like comparing a Ferrari to a BMW.
The SPC benchmarks are representative of typical workloads seen in real-life which means that performance recorded in a benchmark may not actually represent what a customer will actually achieve. Customers are advised to run their own benchmarks to determine real world performance.