To evaluate the efficacy of PMQS and the various load balancing techniques we ran Mkbench and Chat on two different systems, (a) a 16-way NUMA system and (b) an 8-way SMP system. The 16-way NUMA system consists 4 quad building blocks. Each quad consisted of 4x450 MHz PIII processors with 512 KB caches and 1GB main memory. We utilized the NUMA patch of Martin Bligh (IBM LTC) for the Linux 2.4.5 kernel to ensure proper interrupt and I/O routing between quads. This particular NUMA patch does not provide NUMA memory abstraction and treats the memory of all nodes as a flat physical memory space. As such all kernel data structures will be located on the first node.The 8-way SMP system is an IBM Netfinity 8500R with 700MHZ PIII processors, 2MB caches and 2.5GB of main memory. The kernel version used on this machine was 2.4.7.
We study the effect of load, poolsize, and workload nature on the overall performance for the five different load balancing methods LBOFF, IP, LBC, LBP-10 and LBP-45. Since, as seen in section 6 the workload nature is one of the overriding differentiators, we show results for each type of workload separately.