I originally posted this at http://dtrace.org/blogs/brendan/2011/12/08/2000x-performance-win
I recently helped analyze a performance issue in an unexpected but common place, where the fix improved performance of a task by around 2000x (two thousand times faster). As this is short, interesting and useful, I've reproduced it here in a lab environment to share details and screenshots.
Issue
In a production SmartOS cloud environment, a script is used to count entries in an active log. This script is executed from cron(1M), and the grep(1) command it uses is running very slowly indeed:
# time grep done log | wc -l 1492751 real 8m56.062s user 8m12.275s sys 0m0.218s # ls -l log -rw-r--r-- 1 root root 18M 2011-12-08 21:59 log
That took almost nine minutes. The grep(1) command is processing this file at about 34 Kbytes/sec, which seems awfully slow. Dropping the "wc -l" and using "grep -c" instead didn't make much difference.
Hypothesis
I have a method for attacking performance issues (which I'll explain in a later article); it starts with checking for errors, because they do cause performance issues and are a fast area to check.
The slow time sounds like it could be disk I/O related: either slow reads caused by disk errors, or random disk reads. I first checked the error counts:
# iostat -En | grep Hard sd0 Soft Errors: 0 Hard Errors: 0 Transport Errors: 0 sd1 Soft Errors: 0 Hard Errors: 0 Transport Errors: 0
Which were fine, and then for random reads:
# iostat -xnz 1
extended device statistics
r/s w/s kr/s kw/s wait actv wsvc_t asvc_t %w %b device
0.2 0.3 1.3 1.0 0.0 0.0 0.0 0.0 0 0 ramdisk1
12.3 81.6 355.4 3005.9 0.0 0.1 0.0 0.7 0 2 sd1
extended device statistics
r/s w/s kr/s kw/s wait actv wsvc_t asvc_t %w %b device
0.0 7.0 0.0 245.7 0.0 0.0 0.0 0.1 0 0 sd1
extended device statistics
r/s w/s kr/s kw/s wait actv wsvc_t asvc_t %w %b device
extended device statistics
r/s w/s kr/s kw/s wait actv wsvc_t asvc_t %w %b device
0.0 4.0 0.0 59.5 0.0 0.0 0.0 0.1 0 0 sd1
[...]
Which are not present. This doesn't look disk related.
Analysis
A great way to see what threads are doing is with "prstat -mLc 1", which uses thread microstate accounting (only available on Solaris-related systems):
# prstat -mLc 1 [...] PID USERNAME USR SYS TRP TFL DFL LCK SLP LAT VCX ICX SCL SIG PROCESS/LWPID 14566 root 100 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0 13 0 0 app1/1 12740 root 98 0.0 0.0 0.0 0.0 0.0 0.0 2.1 1 26 8 0 grep/1 18998 root 3.0 12 0.0 0.0 0.0 0.0 85 0.0 45 0 5K 0 app2/1 17347 root 0.1 0.2 0.0 0.0 0.0 0.0 100 0.0 16 0 170 0 prstat/1 29693 root 0.0 0.0 0.0 0.0 0.0 0.0 100 0.0 13 0 128 0 sshd/1 12741 root 0.0 0.0 0.0 0.0 0.0 0.0 100 0.0 7 0 7 0 wc/1 [...]
The output shows grep(1) spending 98% of its time in USR (running in user-land), and 2.1% in LAT (CPU dispatcher queue latency - ready to run). There is no high percentage of SLP (sleep) time for disk I/O, which we already suspected from using iostat(1) earlier. (This lab system is also running a couple of dummy applications to simulate production load.)
Hot CPU issues are some of the easiest to solve, usually through profiling. Here, I've used DTrace to sample grep(1) user-land stack traces at 997 Hertz. I ran this one-liner for several seconds, then hit Ctrl-C to stop sampling and print the report:
# dtrace -n 'profile-997 /execname == "grep"/ { @[ustack()] = count(); }'
dtrace: description 'profile-997 ' matched 1 probe
^C
[...]
libc.so.1`_UTF8_mbrtowc
grep`check_multibyte_string+0x85
grep`EGexecute+0xa6
grep`grepbuf+0xf2
grep`grepfile+0x4f3
grep`main+0x1157
grep`_start+0x6c
406
libc.so.1`mbrtowc+0x1d
grep`EGexecute+0xa6
grep`grepbuf+0xf2
grep`grepfile+0x4f3
grep`main+0x1157
grep`_start+0x6c
498
grep`check_multibyte_string+0xd5
grep`EGexecute+0xa6
grep`grepbuf+0xf2
grep`grepfile+0x4f3
grep`main+0x1157
grep`_start+0x6c
1023
This shows codepaths that were sampled. At the top is the on-CPU function, then its ancestory. The number is the count of samples for that stack trace.
The hot functions included check_multibyte_string(), mbrtowc() and _UTF8_mbrtowc(). check_multibyte_string() is declared in grep(1) (prefixed with "grep`"), and mbrtowc() is from libc. I'll start with mbrtowc() - given that it's in libc, it probably has a man page.
# man mbrtowc
mbrtowc(3C) Standard C Library Functions mbrtowc(3C)
NAME
mbrtowc - convert a character to a wide-character code (restartable)
[...]
Wide character conversion?
Hypothesis #2
The team I was working with suggested this could be locale related:
# echo $LANG en_US.UTF-8
I'd also punched "grep mbrtowc" into google and found grep - Bugs: bug #14472, grep is slow in multibyte locales [Savannah]". But that's the GNU version of grep(1), which we didn't think we were using.
# which grep /opt/local/bin/grep # grep --version GNU grep 2.5.3 [...]
Ah. So we are using GNU grep(1), and we have LANG set to en_US.UTF-8.
Experiment & Fix
# time LANG=C grep done log | wc -l 1492751 real 0m0.304s user 0m0.361s sys 0m0.057s
That's over 1700x faster.
And using the SmartOS grep(1):
# time /usr/bin/grep done log | wc -l 1492751 real 0m0.268s user 0m0.318s sys 0m0.059s
That's 2000x faster.
On an idle system, these improvements are between 100x - 200x (still huge). It seems that with production load, the long running grep(1) process takes even longer than usual. I'd suspect this was due in part to thread priorities and the kernel scheduler, giving the longer-running grep(1) a lower priority and less CPU time per-second than the short-lived grep(1). (If I needed to confirm this, I can examine the kernel scheduling events and priorities using DTrace.)
Cause
Both PATH and LANG had changed in the new systems, causing the GNU grep to be executed with a non-C locale. The environment was set is these locations:
# grep LANG= /etc/default/init LANG=en_US.UTF-8 # grep PATH= ~/.profile MANPATH=/opt/local/man:/usr/share/man PATH=/opt/local/bin:/opt/local/sbin:/usr/bin:/usr/sbin # grep PATH= /etc/default/cron PATH=/opt/local/bin:/opt/local/sbin:/usr/bin:/usr/sbin:/usr/sfw/bin SUPATH=/opt/local/bin:/opt/local/sbin:/usr/bin:/usr/sbin:/usr/sfw/bin
Happy grep(1)
In case you are curious, here's what GNU grep is doing when LANG=C, also profiled with DTrace:
[...]
grep`prtext+0x230
grep`grepbuf+0x51
grep`grepfile+0x4f3
grep`main+0x1157
grep`_start+0x6c
7
grep`EGexecute+0x4e5
grep`grepbuf+0xf2
grep`grepfile+0x4f3
grep`main+0x1157
grep`_start+0x6c
7
libc.so.1`mutex_lock_impl+0xf8
libc.so.1`mutex_lock+0xc
libc.so.1`cancel_safe_mutex_lock+0x9
libc.so.1`_flockget+0x1b
libc.so.1`fwrite+0x67
grep`prline+0x39a
grep`prtext+0x21d
grep`grepbuf+0x51
grep`grepfile+0x4f3
grep`main+0x1157
grep`_start+0x6c
15
libc.so.1`clear_lockbyte+0x2c
libc.so.1`mutex_unlock_queue+0x43
libc.so.1`mutex_unlock+0x12c
libc.so.1`cancel_safe_mutex_unlock+0x1f
libc.so.1`fwrite+0xa2
grep`prline+0x39a
grep`prtext+0x21d
grep`grepbuf+0x51
grep`grepfile+0x4f3
grep`main+0x1157
grep`_start+0x6c
18
Much lower sample counts (since it ended faster), and no mbrtowc(). The hottest two stacks show lock activity (which is where I'd start if I wanted to improve GNU grep(1) performance further).
A bit faster
Using LANG=C and grep(1)'s inbuilt "-c" to do counting instead of "wc -l", makes it run even faster. Usually faster than 0.2s, and as fast as 0.12s:
# time LANG=C grep -c done log 1492751 real 0m0.120s user 0m0.110s sys 0m0.010s
Over 4000x.
Conclusion
Picking a non-C locale can hurt performance - something that has been known for many years. In this case, a GNU grep(1) bug inflated the translation overhead to slow down performance by a huge degree: up to 2000x. In short:
Leave LANG=C.
Aim DTrace at everything - even grep(1).