This is the mail archive of the cygwin-developers mailing list for the Cygwin project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Performance optimization in av::fixup - use buffered IO, not mapped file


On 12/12/2012 2:21 PM, Christopher Faylor wrote:
On Wed, Dec 12, 2012 at 12:11:46PM -0500, Ryan Johnson wrote:
On 12/12/2012 12:03 PM, Christopher Faylor wrote:
On Tue, Dec 11, 2012 at 07:13:04PM -0800, Daniel Colascione wrote:
On 12/11/2012 5:06 PM, Daniel Colascione wrote:
On 12/10/2012 7:51 PM, Daniel Colascione wrote:
The key to generating a binary that repros the problem is to unexec emacs, then
try to repro with that generated binary, not a copy of it.
The real explanation is a lot simpler: the binary is sparse. When you create a
file mapping object for a sparse file, Windows discards all cached pages for
that file. It makes sense that compilers (and Emacs unexec) would create sparse
files as they seek around inside their outputs.
Anyway, the binary is sparse because our linker produces sparse files.

Would the Cygwin developers accept this patch? With it, applications would need
to explicitly use ftruncate to make files sparse. Considering the horrible and
unexpected performance implications of sparse files, I don't think generating
them automatically from a sequence of seeks and writes is the right thing to do.
I don't know if this was already done (don't see it in a quick glance at
the archives) but, if this is just a simple case of executable files
being sparse, it seems like an obvious optimization would be to just to
do a, e.g.,

cp --sparse=never -p foo.exe foo.exe.tmp
mv foo.exe.tmp foo.exe

Wouldn't that remove the sparseness and wouldn't you see astounding
performance improvments as a result?
Nope. You'd have to rm foo.exe first.
"mv" does that automatically, doesn't it?
Oops. Thinko. I had "cp" in my head.

Doing so fixes the problem nicely, though, as you suggest.
I don't think we should be considering ripping code out of Cygwin
without some actual data to back up claims.  Testing something like the
above should make it easier to justify.

I'm actually rather surprised that setup.exe's tar code would maintain an
executable's sparseness.
Setup is fine. It's home-brew stuff that suffers, unless/until invoking
`make install' copies the sparse file to its final destination, losing
the sparse property along the way.

Personally, I'm still in shock that the loader barfs so badly over
sparse files... normal reads via mmap and fread use the fs cache just fine.
If we're talking about the loader then why are the examples I asked for
using "cp"?  That's not really apples-to-apples.
I think I'm miscommunicating something here... let me try again.

Problem:

Attempts to execute a file with the NTFS "sparse" attribute force the Windows loader to bypass the fs cache and fetch the file from disk. Long start times result, and are a pain if the executable is invoked frequently. The problem gets worse if you fill in all the holes, because it's still "sparse" but now has more bytes on disk to fetch. This is arguably a bug in Windows.

Steps to repro:

1. Arrange for the creation of a sparse file
2. Execute said file, and enjoy the delay while the loader bypasses the file cache to get the data directly from disk


"cp" with its "--sparse" option is merely an easy way to accomplish step #1. Step #2 is where all the fun happens.

STC attached. Output on my machine is below.

Workaround: copy the file to strip the flag, disable the sparse file optimization in lseek(), or replace lseek/write pairs that write out executable files with calls to pwrite (which apparently lacks the optimization).


$ ./slow.sh 2>&1 | grep '\(real\|sparse\)' non-sparse original runs quickly real 0m0.078s sparse copy can be read quickly from fs cache real 0m0.036s sparse copy slow-to-run in spite of being cached real 0m2.969s sparse copy no longer fully cached real 0m0.911s filling all holes makes sparse penalty worse real 0m5.289s

Ryan

Attachment: slow.sh
Description: Text document


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]