At Wed, 16 Nov 2011 02:59:21 -0500, Christoph Hellwig wrote: > > Di you run any benchmarks and/or look at memory use with an without > tcmalloc? It worked great where I used it, but I really like to have > numers for changes like this. I've run the following benchmark on VM: for bs in 4 16 64 256 1024; do dd if=/dev/zero of=/dev/sdb bs=${bs}k oflag=direct count=1024 done Sheepdog is started with one node and running on ext3 (barrier=1). The result are as follows: bs with tcmalloc without tcmalloc --------------------------------------- 4 KB 488 kB/s 487 kB/s 16 KB 1.9 MB/s 1.9 MB/s 64 KB 7.2 MB/s 7.1 MB/s 256 KB 22.4 MB/s 22.3 MB/s 1 MB 34.4 MB/s 34.4 MB/s There were no obvious difference between them, so for now I think of dropping the tcmalloc patch. Let me know if anyone got different results. Thanks, Kazutaka |