Linux/perl mmap performance

问题:

I'm trying to optimize handling of large datasets using mmap. A dataset is in the gigabyte range. The idea was to mmap the whole file into memory, allowing multiple processes to work on the dataset concurrently (read-only). It isn't working as expected though.

As a simple test I simply mmap the file (using perl's Sys::Mmap module, using the "mmap" sub which I believe maps directly to the underlying C function) and have the process sleep. When doing this, the code spends more than a minute before it returns from the mmap call, despite this test doing nothing - not even a read - from the mmap'ed file.

Guessing, I though maybe linux required the whole file to be read when first mmap'ed, so after the file had been mapped in the first process (while it was sleeping), I invoked a simple test in another process which tried to read the first few megabytes of the file.

Suprisingly, it seems the second process also spends a lot of time before returning from the mmap call, about the same time as mmap'ing the file the first time.

I've made sure that MAP_SHARED is being used and that the process that mapped the file the first time is still active (that it has not terminated, and that the mmap hasn't been unmapped).

I expected a mmapped file would allow me to give multiple worker processes effective random access to the large file, but if every mmap call requires reading the whole file first, it's a bit harder. I haven't tested using long-running processes to see if access is fast after the first delay, but I expected using MAP_SHARED and another separate process would be sufficient.

My theory was that mmap would return more or less immediately, and that linux would load the blocks more or less on-demand, but the behaviour I am seeing is the opposite, indicating it requires reading through the whole file on each call to mmap.

Any idea what I'm doing wrong, or if I've completely misunderstood how mmap is supposed to work?

回答1:

Ok, found the problem. As suspected, neither linux or perl were to blame. To open and access the file I do something like this:

#!/usr/bin/perl
# Create 1 GB file if you do not have one:
# dd if=/dev/urandom of=test.bin bs=1048576 count=1000
use strict; use warnings;
use Sys::Mmap;

open (my $fh, "<test.bin")
    || die "open: $!";

my $t = time;
print STDERR "mmapping.. ";
mmap (my $mh, 0, PROT_READ, MAP_SHARED, $fh)
    || die "mmap: $!";
my $str = unpack ("A1024", substr ($mh, 0, 1024));
print STDERR " ", time-$t, " seconds\nsleeping..";

sleep (60*60);

If you test that code, there are no delays like those I found in my original code, and after creating the minimal sample (always do that, right!) the reason suddenly became obvious.

The error was that I in my code treated the $mh scalar as a handle, something which is light weight and can be moved around easily (read: pass by value). Turns out, it's actually a GB long string, definitively not something you want to move around without creating an explicit reference (perl lingua for a "pointer"/handle value). So if you need to store in in a hash or similar, make sure you store \$mh, and deref it when you need to use it like ${$hash->{mh}}, typically as the first parameter in a substr or similar.

回答2:

If you have a relatively recent version of Perl, you shouldn't be using Sys::Mmap. You should be using PerlIO's mmap layer.

Can you post the code you are using?

回答3:

On 32-bit systems the address space for mmap()s is rather limited (and varies from OS to OS). Be aware of that if you're using multi-gigabyte files and your are only testing on a 64-bit system. (I would have preferred to write this in a comment but I don't have enough reputation points yet)

回答4:

one thing that can help performance is the use of 'madvise(2)'. probably most easily done via Inline::C. 'madvise' lets you tell the kernel what your access pattern will be like (e.g. sequential, random, etc).

回答5:

That does sound surprising. Why not try a pure C version?

Or try your code on a different OS/perl version.

回答6:

See Wide Finder for perl performance with mmap. But there is one big pitfall. If your dataset will be on classical HD and you will read from multiple processes, you can easily fall in random access and your IO will fall down to unacceptable values (20~40 times).

回答7:

Ok, here's another update. Using Sys::Mmap or PerlIO's ":mmap" attribute both works fine in perl, but only up to 2 GB files (the magic 32 bit limit). Once the file is more than 2 GB, the following problems appear:

Using Sys::Mmap and substr for accessing the file, it seems that substr only accepts a 32 bit int for the position parameter, even on systems where perl supports 64 bit. There's at least one bug posted about it:

#62646: Maximum string length with substr

Using open(my $fh, "<:mmap", "bigfile.bin"), once the file is larger than 2 GB, it seems perl will either hang/or insist on reading the whole file on the first read (not sure which, I never ran it long enough to see if it completed), leading to dead slow performance.

I haven't found any workaround to either of these, and I'm currently stuck with slow file (non mmap'ed) operations for working on these files. Unless I find a workaround I may have to implement the processing in C or another higher level language that supports mmap'ing huge files better.

回答8:

If I may plug my own module: I'd advice using File::Map instead of Sys::Mmap. It's much easier to use, and is less crash-prone than Sys::Mmap.

回答9:

Your access to that file had better be well random to justify a full mmap. If your usage isn't evenly distributed, you're probably better off with a seek, read to a freshly malloced area and process that, free, rinse and repeat. And work with chunks of multiples of 4k, say 64k or so.

I once benchmarked a lot string pattern matching algorithms. mmaping the entire file was slow and pointless. Reading to a static 32kish buffer was better, but still not particularly good. Reading to freshly malloced chunk, processing that and then letting it go allows kernel to work wonders under the hood. The difference in speed was enormous, but then again pattern matching is very fast complexitywise and more emphasis must be put on handling efficiency than perhaps is usually needed.