Mandrake 7.0 Hard Drive Optimizations

Oh God, Not Another Benchmark

Yes, Benchmark Brian is back (say that five times fast). While gauging the performance of Tyan's Trinity 400 motherboard, I wondered how much difference those Mandrake 7.0 hard drive optimizations would make. Since I had a few hours to kill, I decided to compare my last Trinity 400 benchmark to an installation of Mandrake with optimizations turned on. This is my system:

• Intel Pentium II 400 MHz (4X100Mhz)
• 64MB PC100 SDRAM
• Seagate 8.4GB UDMA/66 Hard Drive (with 80-pin cable)
• ATI Rage Pro AGP 2X (8 MB)
• Acer 50X ATAPI CD-ROM
• Intel Pro/100 Ethernet

And The Winner Is...

My test arsenal: Mandrake 7.0 (installed as "custom/development" with all 1395MB of packages), BYTEmark v2 and UNIX Bench 4.0.1.. Let's go to the videotape!

UNIX Bench 4.0.1 (index scores) No Opt. With Opt.

• Arithmetic Test (type = double)
  • No Opt.: 33.3
  • With Opt.: 33.3
• Dhrystone 2
  • No Opt.: 73.4
  • With Opt.: 73.4
• Excel Throughput
  • No Opt.: 159.3
  • With Opt.: 157.7
• File Copy (1024 buffsize, 2000 maxblocks)
  • No Opt.: 188.0
  • With Opt.: 186.2
• File Copy (256 buffsize, 500 maxblocks)
  • No Opt.: 217.7
  • With Opt.: 216.7
• File Copy (4096 buffsize, 8000 maxblocks)
  • No Opt.: 5.0
  • With Opt.: 22.4
• Pipe Throughput
  • No Opt.: 186.1
  • With Opt.: 185.6
• Process Creation
  • No Opt.: 255.3
  • With Opt.: 256.5
• Shell Scripts (8 concurrent)
  • No Opt.: 126.7
  • With Opt.: 126.7
• System Call Overhead
  • No Opt.: 148.9
  • With Opt.: 148.3
• Final Score
  • No Opt.: 96.7
  • With Opt.: 112.0

Wow, what a difference a day makes! That increase in UNIX Bench's final score comes from the "File Copy (4096 buffsize, 8000 maxblocks)" test. Note this test is geared for copying large files, like a server might do. I assume Mandrake makes better use of 32-bit transfer and DMA transfer when optimizations are enabled (small file copying doesn't benefit much from DMA transfer).

MandrakeUser has a nice article on optimizing hard drive performance. It describes using hdparm to improve drive performance, and some simple system configuration tips (lots of RAM, no slave drives, short cables, etc.). This article introduced me to using hdparm as a benchmark. The scores are shown below (note a more-than-slight improvement when optimizations are enabled):

hdparm -Tt /dev/hda

• Timing buffer-cache reads (MB/sec)
  • No Opt.: 100.79
  • With Opt.: 102.40
• Timing buffered disk reads (MB/sec)
  • No Opt.: 3.00
  • With Opt.: 16.98

Of course, you should heed Mandrake's warning about the system becoming unstable. DMA transfer shouldn't be used on older 486 and Pentium motherboards, and some hard drives are better than others. The major benefit will be seen working with large files, but some Linux apps are pretty damn big, so this will help with games and a number of commercial applications.

But UDMA/66 with hard disk optimizations should be even better. When Mandrake 7.1 comes out, we'll take that monster on.

Brian Richardson listens to a large collection of MP3z on his wireless headphones while wandering the cubicle filled hallways of AMI in Norcross, Georgia. But don't worry, none of them are by Metallica.