New submission from Linlin Yan:

I noticed this problem when I run a Python2 program (MACS: 
http://liulab.dfci.harvard.edu/MACS/) very inefficiently on a large storage on 
a high performace server (64-bit Linux). It was much slower (more than two 
days) than running it on a normal PC (less than two hours).

After ruling out many optimizing conditions, I finally located the problem on 
the seek() function of Python2. Now I can reproduce the problem in a very 
simple example:

#!/usr/bin/python2
f = open("Input.sort.bam", "rb")
f.seek(0, 2)
f.close()

Here, the size of file 'Input.sort.bam' is 4,110,535,920 bytes. When I run the 
program with 'strace' to see the system calls on Linux:

$ strace python2 foo.py
...
open("Input.sort.bam", O_RDONLY)        = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=4110535920, ...}) = 0
fstat(3, {st_mode=S_IFREG|0644, st_size=4110535920, ...}) = 0
mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 
0x7f23d4492000
fstat(3, {st_mode=S_IFREG|0644, st_size=4110535920, ...}) = 0
lseek(3, 4110532608, SEEK_SET)          = 4110532608
read(3, 
"f\203\337<\334\350\313\315\345&T\227\211\fC\212a\260\204P\235\366\326\353\230\327>\373\361\221\357\373"...,
 3312) = 3312
close(3)                                = 0
...

It seems that python2 just move file cursor to a specific position (4110532608 
in this case) and read ahead the rest bytes, rather than seek to the file end 
directly. I tried to run the exact the same program on the large storage, the 
position changed to 1073741824, left 889310448 bytes to read to reach the file 
end, which reduced the performance a lot!

----------
components: IO
messages: 219586
nosy: yanlinlin82
priority: normal
severity: normal
status: open
title: Seeking to EOF is too inefficient!
type: performance
versions: Python 2.7

_______________________________________
Python tracker <rep...@bugs.python.org>
<http://bugs.python.org/issue21638>
_______________________________________
_______________________________________________
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

Reply via email to