Hi all, I'm using Trafficserver 4.2.0 on Debian Wheezy, with raw disks as cache on VMware virtual machines. Activity level is a bit less than 400 requests per seconds with a cache hit ratio greater than 99%.
On a regular basis, traffic_server process (ET_NET 0) starts to heavily use CPU (more than 75%). That's fine for me since it's the only process on the virtual machine. But, the problem is that most of this CPU usage concerns IO Wait. Running command iostat -m 1 confirms that all IO Wait operations are related to cache raw devices. Restarting trafficserver service fixes the problem... till next time. I tried to strace the process and got plenty of: epoll_wait(21, {{EPOLLIN, {u32=301105424, u64=47154747048208}}}, 32768, 10) = 1 read(14, "\1\0\0\0\0\0\0\0", 8) = 8 futex(0x2ae31800d07c, FUTEX_WAKE_OP_PRIVATE, 1, 1, 0x2ae31800d078, {FUTEX_OP_SET, 0, FUTEX_OP_CMP_GT, 1}) = 1 futex(0x2ae31800d050, FUTEX_WAKE_PRIVATE, 1) = 1 Found nothing strange in diags.log, manager.log or traffic.out log files. Any idea welcome, Regards, Jean-Baptiste Favre