Avoiding system calls

October 1, 2012

GravatarBy Kazu Yamamoto

This is fifth article for the series of "Improving the performance of Warp". Readers are supposed to read the following articles:

  1. Improving the performance of Warp
  2. Sending header and body at once
  3. Caching file descriptors
  4. Composing HTTP response headers

In this article, I will explain how to avoid the fcntl() system call and the gettimeofday() vsyscall.

Avoiding fcntl()

I sometimes compare Mighttpd and nginx with the results of strace on Linux and/or ktrace on BSD variants. One day, I noticed that nginx uses the accept4() system call, which I did not know about at that moment. (I used to be an expert of BSD variants but am a newbie to Linux.)

In the low level of GHC, file/socket operations are basically implemented as non-blocking. If the network package is used, a listening socket is created with the non-blocking flag set. When a new connection is accepted from the listening socket, it is necessary to set the corresponding socket as non-blocking, too. The network package implements this by calling fcntl() twice: one is to get the current flags and the other is to set the flags with the non-blocking flag ORed.

On Linux, the non-block flag of a connected socket is always unset even if its listening socket is non-blocking. The accept4() system call is an extension version of accept() on Linux. It can set the non-blocking flag when accepting. So, if we use accept4(), we can avoid two unnecessary fcntl()s. I modified the network package to use accept4 on Linux and it is included in version 2.3.1.0 or later.

On BSD variants the non-block flag of a connected socket is inherited from its listening socket. So, we can also avoid two unnecessary fcntl()s on BSD variants. But I have not implemented this.

Avoiding gettimeofday()

Date strings are used in various ways. An end HTTP server should return GMT date strings in header fields such as Date:, Last-Modified:, etc:

Date: Mon, 01 Oct 2012 07:38:50 GMT

For logging, a local date string in the Apache style would be convenient:

01/Oct/2012:16:42:33 +0900

It is known that formatTime in the Data.Time.Format module is too slow for high performance servers. Years ago, I implemented faster format packages: http-date for the former and unix-time for the latter. Unfortunately, they are still slow for high performance servers. And if an HTTP server accepts more than one request per second, the server repeats the same formatting again and again. So, formatted date strings should be cached.

The members of the web-devel mailing-list discussed these issues with the following assumption:

  1. Formatting time to date string is a heavy job
  2. Getting the current time by gettimeofday() is a light job

The discussion resulted in the following algorithm:

  • When a formatted date is required, first issue getttimeofday(). Then compare it with a cached time.
  • If they are equal, return the cached formatted date.
  • Otherwise, format the new time to a new formatted date, cache them, and return the new formatted date.

It seems to me that the assumption 2 is not correct. gettimeofday() was a system call in old Linux while it is a vsyscall in new Linux. To my experience, I cannot say that calling gettimeofday() is a light job on both old and new Linux.

So, I implemented a new algorithm:

  • Designated Haskell thread issues gettimeofday() every second, formats the result time to a date, and caches it.
  • When a formatted date is required, the cached formatted date is simply returned.

Only end HTTP servers should return the Date: header field. Since Warp can be used to implement proxies, adding Date: is not warp's job. Also, logging is not warp's job, either. WAI applications should take care of both adding Date: and logging.

Yesod and Mighttpd both use the fast-logger package for logging. I modified it so that it provides both algorithms above. Also, I changed Mighttpd to use the new algorithm to generate Date:. The packages with these modification are already available from Hackage.

Comments

comments powered by Disqus

Archives