VirtuCache Performance Benchmarks Using Iometer
Performance benchmark using Iometer comparing Storage I/O with and without VirtuCache, with multi-threaded 67% read 33% write workload that saturated the disk.
Multi-threaded mixed read-write Iometer tests are a good way to simulate real life OLTP applications for testing both the in-server Flash Card and our Software
I ran Iometer on a relatively small server ( 4core Xeon 3400 with 8GB RAM ). My workload involved simultaneous reads and writes, with a constant proportion of 67% reads and 33% writes spread across 6 threads on a 4 core server. The VM was assigned 4 cores to 1 vCPU, and 4 GB RAM. I used queue depths of 128 and 256 to generate large amounts of load. Windows 2008 Server R2 was installed in a VMware 5.0 VM.
Caching device was a PCIe based Flash Card (costing about $ 8000 per TB) and the data to be cached was on a cheap SCSI disk.
Highlights of the test results are
- Throughput improvements range from 2X to 8X, for both read and write throughput.
- Application response times are reduced by between 30% and 85%.
Comparing these Iometer results to the Sqlio results on my last blog article, it is easily concluded that our caching solution with a high IOPS Flash Card over PCIe improves read performance 40-50X. This performance gain reduces to 2-4X when I increase the write workload to 33% and saturate the disk with a combination of high queue depth (256) and many threads (6 threads on a 4 core CPU).
Iometer charts follow.
Figure 1 : Total read + write throughput (MBps)
Figure 2 : Read throughput in my mixed 67%read+33% write tests (MBps)
Figure 3 : Write throughput in my mixed 67%read+33% write tests (MBps)
Figure 4 : Average response times for both reads and writes (ms)
Figure 5 : Average read response times in my 67%read+33% write tests (ms)
Figure 6 : Average write response times in my 67%read + 33% write tests (ms)