CompoundFileWriter should pre-set its file length

• Key: LUCENE-705
• Type: Improvement
• Status: Closed
• Resolution: Fixed
• Priority: Minor
• Assignee: Michael McCandless
• Reporter: Michael McCandless
• Component/s: Index
• Affects Version/s: 2.1
• Fix Version/s: 2.4

Description

I've read that if you are writing a large file, it's best to pre-set
the size of the file in advance before you write all of its contents.
This in general minimizes fragmentation and improves IO performance
against the file in the future.

I think this makes sense (intuitively) but I haven't done any real
performance testing to verify.

Java has the java.io.File.setLength() method (since 1.2) for this.

We can easily fix CompoundFileWriter to call setLength() on the file
it's writing (and add setLength() method to IndexOutput). The
CompoundFileWriter knows exactly how large its file will be.

Another good thing is: if you are going run out of disk space, then,
the setLength call should fail up front instead of failing when the
compound file is actually written. This has two benefits: first, you
find out sooner that you will run out of disk space, and, second, you
don't fill up the disk down to 0 bytes left (always a frustrating
experience!). Instead you leave what space was available
and throw an IOException.

My one hesitation here is: what if out there there exists a filesystem
that can't handle this call, and it throws an IOException on that
platform? But this is balanced against possible easy-win improvement
in performance.

Does anyone have any feedback / thoughts / experience relevant to
this?

Comments

[Grant Ingersoll]

This seems reasonable, although I am not an expert in low-level file system calls like this. I guess for me the thing would be to find out if the major filesystems support it (Windows, OSX, Linux) and then perhaps we can deal w/ others from there as they arise (i.e. those that don't support it)

[Michael McCandless]

OK I'll test on the major platforms, and take that approach. I'll tentatively target 2.4.

[Michael McCandless]

Attached patch. All tests pass on OS X 10.4, Linux 2.6.22, Windows
Server 2003. I plan to commit in a day or two.

I just committed this.

[Otis Gospodnetic]

Didn't find time to comment on this earlier.
Does this mean one will no longer be able to tell exactly how large the index really is (because some portion of some data files will actually be empty)?

[Michael McCandless]

Does this mean one will no longer be able to tell exactly how large the index really is (because some portion of some data files will actually be empty)?

Only while the CFS is being built. After it's done being built, then
it is "fully occupied" (no portion are empty).