hayamiz / perfmonger Goto Github PK

$ perfmonger server

This command should draw a CPU graph but no graph update because of load failure of /faucet

perfmonger record --kill does not check the background perfmonger-recorder process has exited or not, and fail if the process does not exit correctly.

Also, perfmonger record --kill should fail if there is no background recorder process.

If timeout option is specified, perfmonger-record exits after the specified time has passed.

$ perfmonger record --interval 0.1 --timeout 10 -d sda
{"time": ..., "ioinfo": {... } ... }
(... 10 seconds ...)
== Summary ==
...

Display hostname on HTML5 realtime monitor.

Use a hostname given by command line argument, or try to obtain it automatically.

• record subcommand: gzipped output stream (maybe gzipped file as default for future version?)
• play subcommand: handle gzipped and plain input stream
  • file extension based approach (*.pgr is plain, *pgr.gz is gzipped)
  • transparent file handling approach (check magic number first, then switch plain/gzip read)

• lsblk
• fdisk -l
• LVM(vgscan, lvscan, pvscan)
• hdparm?

add 'include <signal.h>' to perfmonger.h

use File.exist?

• streaming gzip output?
• gzip just after finishing recording session?

Currently summary subcommand reads all records into memory, then do computation. This does not work if records takes memory space larger than physical memory.
Calculate summary in streaming processing fashion.

I want something like this:

$ perfmonger --version
PerfMonger version 0.3.1 (20XX-XX-XX)

Add options to trim start/end duration from recorded logs before calculating summary.

Migrate to Docker-based stack
http://devcenter.wercker.com/docs/migrating-to-workflows-from-classic

procfs ( /proc/* ) is not available on OSX.

sysctl interface seems to be useful? objective c - Get CPU percent usage - Stack Overflow

Summary takes >100 seconds for large perfmonger log files.
Show progress bar for such files.

In a server with a lot of drives, fingerprint subcommand takes so long time with low CPU usage. There must exist some opportunities for speed up.

perfmonger-record starts recording just after a trigger condition is fired.

trigger can be:

• the first launch of the specified program
• the first fork from the specified process
• specified time

or something elese.

Use ANSI color or something like that.

• x[i] ... i-th sample value of IOPS
• t[i] ... beginning time of i-th sample
• N ... # of samples

Currently avg. IOPS is calculated as a average value of sampled IOPS.

avg IOPS = (x[0] + x[1] + ... + x[N-1]) / N

But actually, avg. IOPS needs to be calculated by this way.

avg IOPS = (x[1] * (t[1] - t[0]) + x[2] * (t[2] - t[1]) + ...  + x[N-1] * (t[N-1] - t[N-2])) / (t[N-1] - t[0])

If /dev/mapper/VolGroup00-LogVol00 or /dev/VolGroup00/LogVol00 is passed to perfmonger record via -d option, convert it to corresponding device name dm-x.

This is done by following steps:

1. Build device name dictionary by reading /proc/partitions
2. Parse input and identify the path of logical volume
3. stat(3) the volume and get major/minor number of the volume
4. Look up the dictionary build in step 1. with major/minor number