Torrey Hoffman wrote:
> 
> I've been putting together a new workstation, and in the process of
> building and testing it I ran into an odd performance anomaly.  It
> probably only matters for people benchmarking different reiserfs
> setups, but I'd like to know what causes it.
> 
> In short: If you make a new reiserfs partition, immediately mount it,
> and run benchmarks, it is noticeably slower than the exact same
> partition after a reboot and remount.  There can be as much as a
> 25 % speed increase just from rebooting!
> 
> I've seen this repeatedly on both single disks and a variety of RAID
> setups. Details of my setup and some of the testing I've done are below.
> 
> Note that mkreiserfs warns that rebooting after fdisk is required.
> I haven't been using fdisk at all - I am running mkreiserfs on raw
> drives and raid devices with no partition tables.
> 
> Perhaps the warning should be that a reboot is required after
> running mkreiserfs?
> 
> Anyway, some details:
> 
> My computer setup:
>  - Dual P3 800 (133Mhz) Tyan Tiger MB with VIA chipset, 512 MB RAM
>  - Linux 2.4.9-ac17, no other patches.
>  - Distribution is approximately Mandrake 8.1 (8.0 + most of Cooker)
>  - all benchmark numbers are based on three consecutive runs of
>    "dbench -c ../client.txt 32", reporting the "Throughput" numbers.
> 
>  - All testing was done with no X running - just basic system daemons
>  - System runs from /dev/hda which is an IBM 7200 RPM drive using the
>    onboard VIA IDE controller.
> 
>  - /dev/hd[e,f,g,h] are four 60 GB 5400 RPM Maxtor drives on a Promise
>    Ultra TX2-100, two drives per cable, ATA 100.
>  - init scripts do "hdparm -c1 -d1 -m16 -W1 -A1" for all hard drives
> 
> Raw device speed:
> hdparm -t -T /dev/hde  : buffer-cache 182 MB/s,         disk    = 27.7 MB/s
> 
> Testing on a single disk, /dev/hde, no partition table
> - mkreiserfs /dev/hde, mount, 3 runs of dbench 32               = 36 - 44
> MB/s
> - after reboot, mount, 3 runs of dbench 32                      = 52 - 56
> MB/s
> 
> raid 0 with two IDE masters only (/dev/hde & /dev/hdg), chunk size 2048
> - mkraid, mkreiserfs /dev/md0, mount, 3 runs of dbench 32       = 52 - 57
> MB/s
> - after reboot, mount /dev/md0, 3 runs of dbench 32             = 68 - 69
> MB/s
> 
> After those tests I always rebooted.
> 
> raid 0 - all four Maxtors, chunk size 2048
> - mkraid, mkreiserfs, reboot, mount, 3 dbench runs              = 71 - 73
> MB/s
> 
> raid 0 - all four Maxtors, chunk size 1024
> - mkraid, mkreiserfs, reboot, mount, 3 dbench runs              = 72 - 74
> MB/s
> 
> This machine will eventually have 1 GB of RAM, two Promise Ultra TX2 100
> controllers with one Maxtor drive per cable, and will be running in a RAID 5
> 
> configuration.  For now I'm just experimenting to check speed and stability.
> 
> Torrey Hoffman
Would results be explained by full write cache creating IO demands on subsequent
running of benchmark?

Hans

Reply via email to