Hi, I'm profiling memcpy and seeing strange behavior (for me at least) and wanted to see if some one has an idea what may be happening.
My set up is as follows: I have a Ubuntu 12.04 Linux host running 3.2.0-23 kernel. It has a four 10-core dual-hyper-threaded CPU with 128GB RAM. I have instantiated a VM running same OS as the host with KVM enabled. VM is provided with 16 vCPUs with 6GB guest memory. The QEMU ( qemu-2.3.0-rc3) process is bound on vCPUs 1-16 (included) using taskset. (The idea is each lwp that QEMU spins per vCPU runs on separate hyper-thread) taskset -pc 1 16 $qemupid An application on this VM mallocs two 1GB chunks to be used as source and destination (virtual address at 256 byte boundary). This application spins of *n* threads each performing memcpy in parallel. Each thread is again bound to a vCPU using *pthread_setaffinity_np()* inside the guest application in a round-robin fashion. The size of each memcpy is 32 MB. I experimented with 4 .. 32 threads in increments of 4. Each thread works on different slice, where slice size is 32 MB. src = bufaligned1 + (slice_sz * j); dst = bufaligned2 + (slice_sz * j); I noticed as the number of thread increases, the time taken to perform memcpy of 32 MB increased too. Time measured in milli-seconds is as given below (Threads - Transfer time) 04 - 52 08 - 79 12 - 148 16 - 180 20 - 223 24 - 270 28 - 302 32 - 354 I was expecting to see the transfer time to be approximately same up until 16 threads as I have 16 vCPUs bound to 16 hyper threads. After 16 threads the time to transfer should increase as multiple threads are being scheduled on the same vCPU and there is resource contention. But why is transfer time increasing between 4-16 threads too ? Looks like there is some contention at the host level. Any ideas what that could be so that I can focus on that and profile that component ? Thanks Shesha
