Callandor

At my work, I have both a Sun T5220 and M3000. The T5220 has 64 ~1.2 GHz “threads” composed of one CPU with 8 cores of 8 threads each. The M3000 has 8 ~2.75GHz threads composed of 4 cores of 2 threads each.

Product installs were taking much longer than usual, and my customer asked me to find out why. The answer: his install is not very parallel, so being limited to a single slower thread really killed the uncompress and install time.

Here are some numbers. Tested on Solaris 10, with files in /tmp which is mounted on RAM. Input file is Sun Studio 12. The .tar file was removed in between tests, and de/compressions of the files were performed before timing them. Tests were repeated once :).

$ ls -l studio12u1-sol-sparc.tar*
-rw-r--r--   1 alucard  1000     530489344 Nov 19 14:57 studio12u1-sol-sparc.tar
-rw-r--r--   1 alucard  1000     523107141 Nov 19 14:57 studio12u1-sol-sparc.tar.bz2
-rw-r--r--   1 alucard  1000     522070280 Nov 19 15:15 studio12u1-sol-sparc.tar.gz

M3000:

$ time gunzip studio12u1-sol-sparc.tar.gz
real    0m6.210s
user    0m5.041s
sys     0m1.168s
$ time bunzip2 studio12u1-sol-sparc.tar.bz2
real    2m13.908s
user    2m11.729s
sys     0m2.110s

T5220:

$ time gunzip studio12u1-sol-sparc.tar.gz 
real    0m33.612s
user    0m30.704s
sys     0m2.905s
$ time bunzip2 studio12u1-sol-sparc.tar.bz2
real    6m29.298s
user    6m25.302s
sys     0m3.985s

The M3000 gunzip’d in 20% of the time the T5220 did, and bunzip’d in 33% of the time. Expected, once you look, because a single thread is much slower on one than on the other.

Now here’s the T5220 with parallel un/compression tools, pigz and pbzip2 (available for Solaris from SunFreeware):

$ time unpigz studio12u1-sol-sparc.tar.gz 
real    0m8.400s
user    0m9.779s
sys     0m3.875s
$ time pbunzip2 studio12u1-sol-sparc.tar.bz2 
real    0m32.492s
user    8m23.074s
sys     0m4.807s

Parallel gunzip completed in 25% of the time as single-threaded on the T5220, and parallel bunzip2 completed in 10% of the time! Amazing what throwing 63 more threads on the task can do!

So if you’re wondering why your Sun T series is slow – try doing more things in parallel :).

Note: to take advantage of pbunzip2’s parallel decompression, the files must be compressed with pbzip2 (said files can still be decompressed with plain bzip2).