Messages in this thread | | | Date | Wed, 6 May 2015 12:10:04 +0800 | From | Yunlong Song <> | Subject | Re: [Question] How does perf still record the stack of a specified pid even when that process is interrupted and CPU is scheduled to other process |
| |
On 2015/4/25 23:53, David Ahern wrote: > On 4/25/15 8:05 AM, Yunlong Song wrote: >> On 2015/4/24 21:58, David Ahern wrote: >>> On 4/24/15 7:31 AM, Yunlong Song wrote: >>>> Now we are profiling the performance of ext4 and f2fs on an eMMC card with iozone, >>>> we find a case that ext4 is better than f2fs in random write under the test of >>>> "iozone -s 262144 -r 64 -i 0 -i 2". We want to analyze the I/O delay of the two >>>> file systems. We have got a conclusion that 1% of sys_write takes up 60% time of >>>> the overall sys_write (262144/64=4096). We want to find out the call stack during >>>> this specific 1% sys_write. Our idea is to record the stack in a certain time period >>>> and since the specific 1% case takes up 60% time, the total number of records of its >>>> stack should also takes up 60% of the total records, then we can recognize those stacks >>>> and figure out what the I/O stack of f2fs is doing in the 1% case. >>> >>> And to address this specific profiling problem have you tried: >>> >>> perf trace record -- iozone ... >>> perf trace -i perf.data -S >>> >>> >>> >>> >> >> But this only shows the system call like strace, but we want the call stack of kernel functions >> in fact. >> > > We haven't added the callchain option yet; on the to-do list. > > perf trace record -g -- iozone ... > perf trace -i perf.data -s > --> summary of system calls, max/min/average times > > perf trace -i perf.data --duration 10.0 -T > --> note the timestamp where the write took a "long" time > > perf script > --> search down to *around* the time of interest; you want the syscall entry; timestamp is for exit > > . >
Hi, David,
It's almost what we want, we are eager to see it can work as a callchain option, since it's really a useful tool in analyzing latency of I/O performance in production case.
-- Thanks, Yunlong Song
| |