Date: Wed, 22 Mar 1995 20:05:58 +1000 From: Bruce Evans <bde@zeta.org.au> To: hackers@FreeBSD.org, phk@ref.tfs.com Cc: faq@FreeBSD.org Subject: Re: Why IDE is bad Message-ID: <199503221005.UAA09696@godzilla.zeta.org.au>
next in thread | raw e-mail | index | archive | help
>I just made a simple test, this shows why IDE is inferior to SCSI for >FreeBSD: Not a fair test. >Western Digital 540 Caviar EIDE disk on IDE controller: >------------------------------------------------------- >dd if=/dev/rwd0d of=/dev/null bs=64k count=8253 > 8253+0 records in > 8253+0 records out > 540868608 bytes transferred in 240 secs (2253619 bytes/sec) >Here is what "systat" tells me: > 2.0%Sys 82.0%Intr 0.5%User 0.0%Nice 15.6%Idle > | | | | | | | | | | > =+++++++++++++++++++++++++++++++++++++++++ When the driver fully supports EIDE, it will use a faster PIO mode at least for reasonable drives/controllers/buses (not non-VLB ISA). This should reduce the largest component of the overhead (transferring data) by a factor of (new PIO speed)/(old PIO speed), typically 11.3/3.3. >Quantum Empire 2100 SCSI-II disk on VL-buslogic controller: >------------------------------------------------------------- >dd if=/dev/rsd0d of=/dev/null bs=64k count=8253 > 8253+0 records in > 8253+0 records out > 540868608 bytes transferred in 110 secs (4916987 bytes/sec) >Here is what "systat" tells me: > 5.6%Sys 20.6%Intr 0.2%User 0.0%Nice 73.6%Idle > | | | | | | | | | | > ===++++++++++ This doesn't show DMA time (except where it increases the Sys and Intr time). The VLB-Buslogic DMA transfer time should be smaller than the IDE transfer time, however. Buslogic claims 40MB/sec bursts while EIDE currently doesn't go above 16.6MB/sec. Your SCSI Intr time seems to be too high. I see only 2.2%Sys 1.4%Intr 0.3%User 0.0%Nice 96.1%Idle for a 486DX2/66 VLB-Buslogic Toshiba-MK537FB transferring at 2.6MB/sec and about twice as much Sys and Intr time for an Quantum-XP34301 transferring at 5.6MB/sec. To find the actual overhead, run a hog process, or better your normal application mix in the background and see how much slower they run. Counting to 100000000 takes 7.61 seconds here while the system is otherwise idle and 10.61 seconds while the Quantum is transferring, so the actual overhead is abount 28%. >CONCLUSION: >----------- >The SCSI-controller transfers twice as much data in the same time, and >uses only 25% of the CPU resources the IDE-controller needs. The better transfer rate is probably because the drive is better. When the driver can transfer at 11.3MB/sec it will be possible to match the SCSI transfer rate at a cost of about 60% overhead - twice as high as for the VLB-Buslogic Quantum-XP34301 combination. In practice, the FreeBSD-current 486DX2/66 VLB-Buslogic Quantum-XP34301 combination spends about 0.01% of its time transferring at >5MB/sec. It goes fast for iozone but very slowly for real files. E.g., `cp -pR /usr/src/sys /usr/src/sys~' takes 85 seconds to copy 2 * 12475K (293K/sec). Any IDE drive can compete with this transfer rate. Bruce
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199503221005.UAA09696>