Linux vs. SCO Benchmarking & Comparison

I had the opportunity to run some benchmarks on identical hardware recently in order to compare the operating system I like (LINUX) with the operating system I use for work (SCO Openserver 5.05).

One of the things I wanted to know was who was faster for disk performance: Linux or Openserver. This is very important to me because the majority of my work is disk-intensive database work. The winner is clearly Linux.

I had suspected this, because occasionally people talk about Linux vs NT and Linux wins, but no one really mentions Openserver against anything, so I didn't expect stellar results from this benchmarking.

The system is: Compaq ML350, Pentium III-600EB with 256k L2 cache on-die at 600Mhz. 512M of 133Mhz ECC SDRAM memory on a ServerWorks/RCC chipset based motherboard. The system had a Compaq/SymbiosLogic/LSI 53x876 dual channel Ultra2 controller. I had three Fujitsu 9G 7200RPM Ultra2 drives on channel 0, tape drive on channel 1. Benchmarking was done on drives 2 and 3 which were completed empty at the time.

Tested: Redhat Linux 6.2, basic install, with a variety of services like httpd running, but no traffic on them, vs SCO Openserver 5.05 with RS505A patches.

In these tests, I used bonnie (not bonnie++, btw) and had it writing out 512 and 1024M data files to single disks.

And note that these are single-threaded benchmarks: they don't really simulate multi-user activity that well, something I hope to test in 2-3 weeks on another machine.

Here's my overview of the results:

1. Linux is faster. Much faster. The columns you really want to pay attention to are the block read and block write columns. Programs doing character i/o are never going to get good performance, so just ignore those columns for the most part.

2. SCO may or may not be slowed down by the fact that it uses a journaled file system. I've used SCO's HTFS filesystem for years and grown spoiled on it; it's very reliable and you don't have to do any fsck's to check the disk for errors after crashes and power outages. In order to tell if SCO was being slowed down by HTFS's journalling overhead, I ran the same benchmarks on the SCO installation using an older, non-journaled file system called EAFS. Unfortunately, this isn't really apples to apples: EAFS is limited to 2G partitions and probably doesn't have many of the performance enhancements of Linux' ext2 or SCO's HTFS. So it was in fact, slower. In order to tell if journalling causes a significant performance hit, we'll just have to wait for a chance to test a journalling filesystem like reiserfs on Linux.

The results:

Compaq ML350 with Redhat Linux 6.2: (should be 4096 byte blocks, btw)

            -------Sequential Output-------- ---Sequential Input-- --Random-- 
            -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
 Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU 

1024        9403  94.5 19004  9.1 8013  11.4  9792 93.6 17177  7.6 142.2 1.4 
1024        8755  87.8 18689  9.3 8189  12.0  9811 93.8 17355  8.0 135.8 1.5 
1024        8706  87.3 16404  8.0 7789  10.7 10019 95.6 17388  8.1 139.1 1.5 
1024        8673  87.0 18053  8.8 8069  10.6  9776 93.3 17307  7.9 139.6 1.4 
1024        9009  90.4 18483  9.2 8190  12.1  9794 93.4 17441  8.0 139.5 1.5 
1024        9094  91.2 17927  8.9 7918  10.5  9941 95.0 17388  7.6 141.5 0.8 
1024        8910  89.3 18116  8.9 7893  10.9  9966 95.2 17499  7.5 137.6 1.3 

Compaq ML350 with SCO Openserver Unix 5.05 (default block size: 1k)

           -------Sequential Output-------- ---Sequential Input-- --Random--
           -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
1024       4571  19.5  6589   9.6 3735  8.3 14998 93.1 21553 26.5 46.7  1.4
1024       4631  20.0  6467   9.4 3742  8.2 14567 90.7 21566 25.8 46.4  1.3
1024       4607  19.2  6604  10.0 3750  8.5 14431 89.5 21571 25.8 46.4  1.2
1024       4500  19.9  6575   9.5 3714  7.9 14651 90.8 21562 26.0 46.5  1.6
1024       4586  19.4  6557   9.3 3722  8.1 14951 92.8 21571 25.8 46.9  1.3
1024       4599  19.8  6452   9.6 3762  8.2 14723 91.5 21544 27.9 46.9  1.5
1024       4596  20.0  6452   9.4 3761  8.2 14469 89.8 21562 25.8 46.7  1.5
1024       4634  20.4  6453   9.3 3767  8.4 14873 92.1 21566 26.4 46.7  1.1
1024       4652  20.4  6434   9.3 3747  8.2 14992 93.1 21571 25.2 46.3  1.3

Ok, so both systems use about the same amount of CPU time to write data, but Linux puts out THREE TIMES AS MUCH DATA. WOW: SCO averages 6.5M per sec while Linux is doing 17-18M/s.

The block input (read) column shows SCO as being faster, but notice how much CPU time it's using to do that? I think SCO is benefiting from the fact that it has a static 6 megabyte disk cache, vs the very dynamic and growing disk cache in Linux.

I have no idea what's allowing SCO to be a bit faster but at the same time requiring it to eat up so much CPU time. I would have guessed Linux to use more CPU time and therefore be a bit slower, but Linux is again being very efficient with the CPU.

The Random seeks column shows Linux and SCO using pretty much the same amount of CPU power, but Linux performing, again, nearly three times as good.

Compaq ML350 SCO Openserver Unix: EAFS (Non-journaled file system)

           -------Sequential Output-------- ---Sequential Input-- --Random--
           -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
1024        4692 20.4  5383 22.8  3748 8.3   6679 41.5  7489  9.2 45.4  1.2
1024        4619 19.9  5418 23.9  3763 8.2   6681 41.5  7489  9.0 44.5  1.4
1024        4621 20.0  5455 23.9  3770 8.3   6622 41.1  7489  9.0 45.1  1.1
1024        4582 19.6  5350 23.4  3799 8.3   6672 41.6  7490  9.2 45.2  1.5
1024        4609 19.8  5060 22.6  3788 8.5   6636 41.3  7490  8.8 44.9  1.0
1024        4596 19.5  5342 23.5  3800 8.1   6619 41.2  7489  8.9 44.9  1.4
1024        4606 19.7  5363 23.1  3775 7.8   6652 41.4  7489  8.7 45.0  1.1
1024        4629 20.3  5414 24.1  3760 8.3   6630 41.2  7489  9.0 45.0  1.2
1024        4568 19.6  5309 24.0  3793 8.3   6668 41.5  7490  9.0 44.8  1.4

Remember, this is SCO with the non-journaled file system. Performance is AWFUL! AWFUL! AWFUL! High CPU overhead and low performance: the worst situation you can imagine other than disk corruption on a regular basis :)

Way to go, Linux!