Messages in this thread | | | Date | Fri, 12 Nov 1999 14:24:06 +0200 | From | Constantine Gavrilov <> | Subject | Mean 2.3.27 kernel |
| |
Hardware: Supermicro P6DGS + 2 PIII 500 MHZ + 1024 MB RAM onboard AIC7895 dual channel SCSI 4 IBM DNES-309170 9GB drives (3 external + 1 internal connected to the same channel). Quantum 10Gb IDE disk.
Software: RedHat 6.1 (stock) + different test custom built kernels (system is installed on a IDE disk).
I wanted to compare IO in 2.2.x and 2.3.x.
Under 2.2.x, I first did: date;dd if=/dev/sda1 of=/dev/sdb1 bs=1024k;date This gives me approximately 11Mbytes/sec.
Then I did, dd `date;if=/dev/sda1 of=/dev/sdb1 bs=1024k;date' and `date;dd if=/dev/sdc1 of=/dev/sdd1 bs=1024k;date' in parallel. This gave me approximately 8Mbytes/sec for each drive (16 Mbytes/sec total).
This is true both for UP and SMP, 2.2.7, 2.2.12 and 2.2.14 kernels.
I wanted to see if 2.3.x would scale better. I compiled 2.3.27 with gcc (not egcs) compiler (on a RedHat 5.2 machine) with IDE, SCSI, and AIC7xxx drivers built in and SMP, rtc and mtrr enabled. I booted multi-user, started gnome, xosview, and two terminals. In one terminal I run date;if=/dev/sda1 of=/dev/sdb1 bs=1024k;date and in the other date;dd if=/dev/sdc1 of=/dev/sdd1 bs=1024k;date
Watching xosview window, I saw that disk meter jumped quickly to above 30MB/sec, than in, a matter of seconds, to zero. The CPU utilization jumped up to 100% on each CPU, all 100% being system time. I waited for more than an hour (on 2.2.x it takes less than 5 minutes). There were no crashes, no lockups, no messages on the console or in the logs. The system was not locked up. Then I hit cntrl-c in both terminals. The dd output let me know that 50% of the job was done (according to the block count). Then I was shocked to see that both my system time and system clock has jumped backwards 8 hours! I rebooted, entered the BIOS and saw that it was true!!! The clock has warped. I am not aware if Cntrl-C caused the time warp. I do know, that at least during the fisrt 15 minutes, there were no visible problems with the clock.
Additional test: I entered BIOS, corrected the clock, and set power management to "on" but disabled all power management features (see note 1 on why). I booted multi-user again, started X in the same fashion, but run only one dd:
date;dd if=/dev/sda1 of=/dev/sdb1 bs=1024k;date
It took the system more than 30 minutes to finish! Again, I observed a very fast (matter of seconds) drop in disk IO and very fast growth of system time. This time, however, the system time was 100% on either one of two CPUs for the first 30-40 seconds, and then both CPUs were saturated. There was no problem with the system or hardware clock this time.
In both test1 and test2 cat `/proc/scsi/aic7xxx/0' often returns random parts of system memory with weird binary simbols, some C code, or even README files from somewhere (I was getting KDE html and db text documentation, for example)! This is especially true during the execution of dd commands.
Additional notes: 1) 2.2.x kernels will detect 64MB of memory on this machine when power management is disabled in BIOS. Enabling power management in BIOS (and disabling all possible power management features) will cause 1GB (970MB) detection. The system works fine in either case (with 64Mb reported and PM disabled or 970MB reported and PM enabled) and passes all of my crash tests. This is the first SuperMicro board I tested that has such behavior (I have done extensive testing of three other models, including a very close P6DBS model currently used as our departmental server. This is unlikely to be a defective board, because I tried two different ones. Bios upgrade does not solve this problem either). 2) 2.3.27 kernel will report 900MB in either case (PM enabled or disabled in BIOS). Is it a correct behavior (I have 1024MB and "high memory support" was disabled)? 3) My hardware clock is set to GMT. 4) ACPI in the BIOS was always disabled. 5) Kernel had no support for PM or ACPI compiled in. 6) During the tests, the X went into screensaver mode several times.
Are these known problems? If anybody (who works on the code) is interested, I can do any sort of additional testing, trying to repeat or refute these results or do whatever else.
-- ---------------------------------------- Constantine Gavrilov Unix System Administrator and Programmer Orbotech Yavne 81102, Israel Phone: (972-8)-942-3064 Fax: (972-8)-942-3800 ----------------------------------------
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |