Clojure

Improve writeClassFile performance

Details

  • Type: Enhancement Enhancement
  • Status: Open Open
  • Priority: Minor Minor
  • Resolution: Unresolved
  • Affects Version/s: Release 1.3
  • Fix Version/s: None
  • Component/s: None
  • Labels:
    None
  • Patch:
    Code

Description

This Discussion about timing issues when writing class files led to the the current implementation of synchronous writes to disk. This leads to bad performance/system-load when compiling Clojure code.

This Discussion questioned the current implentation.

Synchronous writes are not necessary and also do not ensure (crash while calling write) valid classfiles.

These Patches (0001 is just a code cleanup for creating the directory structure) ensures atomic creation of classfiles by using File.renameTo()

Activity

Ascending order - Click to sort in descending order
Hide
Permalink
David Powell added a comment -

Removing sync makes clojure build much faster. I wonder why it was added in the first place? I guess only Rich knows? I assume that it is not necessary.

If we are removing sync though, I wouldn't bother with the atomic rename stuff. Doing that sort of thing can cause problems on some platforms, eg with search indexers and virus checkers momentarily locking files after they are created.

The patch seems to be assuming that sync is there for some reason, but my initial assumption would be that sync isn't necessary - perhaps it was working around some issue that no longer exists?

Show
David Powell added a comment - Removing sync makes clojure build much faster. I wonder why it was added in the first place? I guess only Rich knows? I assume that it is not necessary. If we are removing sync though, I wouldn't bother with the atomic rename stuff. Doing that sort of thing can cause problems on some platforms, eg with search indexers and virus checkers momentarily locking files after they are created. The patch seems to be assuming that sync is there for some reason, but my initial assumption would be that sync isn't necessary - perhaps it was working around some issue that no longer exists?
Hide
Permalink
Jürgen Hötzel added a comment -

Although its unlikely: there is a possible race condition "loading a paritally written classfile"?:

https://github.com/clojure/clojure/blob/master/src/jvm/clojure/lang/RT.java#L393

Show
Jürgen Hötzel added a comment - Although its unlikely: there is a possible race condition "loading a paritally written classfile"?: https://github.com/clojure/clojure/blob/master/src/jvm/clojure/lang/RT.java#L393
Hide
Permalink
John Szakmeister added a comment -

The new improve-writeclassfile-perf version of the patch combines the two previous patches into a single patch file and brings them up-to-date with master. I can split the two changes back out into separate patch files if desired, but I figured out current tooling is more geared towards a single patch being applied.

Show
John Szakmeister added a comment - The new improve-writeclassfile-perf version of the patch combines the two previous patches into a single patch file and brings them up-to-date with master. I can split the two changes back out into separate patch files if desired, but I figured out current tooling is more geared towards a single patch being applied.
Hide
Permalink
John Szakmeister added a comment -

FWIW, both fixes look sane. The first one is a nice cleanup. The second one is a little more interesting in that uses a rename operation to put the final file into place. It removes the sync call, which does make things faster. In general, if we're concerned about on-disk consistency, we should really have a combination of the two: write the full contents to a tmp file, sync it, and atomically rename to the destination file name.

Neither the current master, nor the current patch will guarantee on-disk consistency across a machine wide crash. The current master could crash before the sync() occurs, leaving the file in an inconsistent state. With the patch, the OS may not get the file from file cache to disk before an OS level crash occurs, and leave the file in an inconsistent state as well. The benefit of the patch version is that the whole file does atomically come into view at once. It does have a nasty side effect of leaving around a temp file if the compiler crashes just before the rename though.

Perhaps a little more work to catch an exception and clean up is in order? In general, I like the patched version better.

Show
John Szakmeister added a comment - FWIW, both fixes look sane. The first one is a nice cleanup. The second one is a little more interesting in that uses a rename operation to put the final file into place. It removes the sync call, which does make things faster. In general, if we're concerned about on-disk consistency, we should really have a combination of the two: write the full contents to a tmp file, sync it, and atomically rename to the destination file name. Neither the current master, nor the current patch will guarantee on-disk consistency across a machine wide crash. The current master could crash before the sync() occurs, leaving the file in an inconsistent state. With the patch, the OS may not get the file from file cache to disk before an OS level crash occurs, and leave the file in an inconsistent state as well. The benefit of the patch version is that the whole file does atomically come into view at once. It does have a nasty side effect of leaving around a temp file if the compiler crashes just before the rename though. Perhaps a little more work to catch an exception and clean up is in order? In general, I like the patched version better.
Hide
Permalink
Ivan Kozik added a comment - - edited

File.renameTo returns false on (most?) errors, but the patch doesn't check for failure. Docs say "The return value should always be checked to make sure that the rename operation was successful." Failure might be especially likely on Windows, where files are opened by others without FILE_SHARE_DELETE.

Show
Ivan Kozik added a comment - - edited File.renameTo returns false on (most?) errors, but the patch doesn't check for failure. Docs say "The return value should always be checked to make sure that the rename operation was successful." Failure might be especially likely on Windows, where files are opened by others without FILE_SHARE_DELETE.

People

Vote (0)
Watch (4)

Dates

  • Created:
    Updated:
Atlassian JIRA (v4.4#649-r158309) | Report a problem
Powered by a free Atlassian JIRA open source license for Clojure. Try JIRA - bug tracking software for your team.