linux-perf

This library is for parsing, representing in Haskell and pretty printing the data file output of the Linux perf command. The perf command provides performance profiling information for applications running under the Linux operating system. This information includes hardware performance counters and kernel tracepoints.

Modern CPUs can provide information about the runtime behaviour of software through so-called hardware performance counters http://en.wikipedia.org/wiki/Hardware_performance_counter. Recent versions of the Linux kernel (since 2.6.31) provide a generic interface to low-level events for running processes. This includes access to hardware counters but also a wide array of software events such as page faults, scheduling activity and system calls. A userspace tool called perf is built on top of the kernel interface, which provides a convenient way to record and view events for running processes.

The perf tool has many sub-commands which do a variety of things, but in general it has two main purposes:

1. Recording events.

2. Displaying events.

The perf record command records information about performance events in a file called (by default) perf.data. It is a binary file format which is basically a memory dump of the data structures used to record event information. The file has two main parts:

1. A header which describes the layout of information in the file (section sizes, etcetera) and common information about events in the second part of the file (an encoding of event types and their names).

2. The payload of the file which is a sequence of event records.

Each event field has a header which says what general type of event it is plus information about the size of its body.

There are nine types of event:

1. PERF_RECORD_MMAP: memory map event.

2. PERF_RECORD_LOST: an unknown event.

3. PERF_RECORD_COMM: maps a command name string to a process and thread ID.

4. PERF_RECORD_EXIT: process exit.

5. PERF_RECORD_THROTTLE:

6. PERF_RECORD_UNTHROTTLE:

7. PERF_RECORD_FORK: process creation.

8. PERF_RECORD_READ:

9. PERF_RECORD_SAMPLE: a sample of an actual hardware counter or a software event.

The PERF_RECORD_SAMPLE events (samples) are the most interesting ones in terms of program profiling. The other events seem to be mostly useful for keeping track of process technicalities. Samples are timestamped with an unsigned 64 bit word, which records elapsed nanoseconds since some point in time (system running time, based on the kernel scheduler clock). Samples have themselves a type which is defined in the file header and linked to the sample by an integer identifier.

Below is an example program which reads a perf.data file and prints out the number of events that it contains.

module Main where

import Profiling.Linux.Perf (readPerfData)
import Profiling.Linux.Perf.Types (PerfData (..))
import System.Environment (getArgs)

main :: IO ()
main = do
   args <- getArgs
   case args of
      [] -> return ()
      (file:_) -> do
         perfData <- readPerfData file
         print $ length $ perfData_events perfData