20040427: Linux RAID revisited (cont.)

>From: Tom Yoksas <address@hidden>
>Organization: UCAR/Unidata
>Keywords:  200404271436.i3REaVCT020136 Linux RAID LDM

Hi Gerry,

We have been looking at the output from 'dmesg' on bigbird to see
if there is anything in your setup that jumps out as being a potential
performance problem.  I wanted to do this since a dual Athlon 2800+
machine here in my office is ingesting and decoding ALL IDD feeds
into GEMPAK format and most into McIDAS format and it does not
show the high load averages that we continually see on bigbird.
For reference, my machine has:

2x Athlon 2800+
Promise Technologies TX2 Ultra 133 IDE controller (for 66 Mhz PCI)
500 GB software RAID (2x Maxtor 7Y250P0 HDs)
1 GB PC2100 ECC Reg RAM
80 GB system HD (Samsung SP0802N)
Tyan Tiger S2466-4M motherboard
Jaton GEFORCE4 Ti4200 128 MB 8X AGP video
3COM 3C509C Ethernet (10/100) - one on motherboard; one extra
NEC ND-2500A 8x DVD+/-RW
Mitsumi 3.5" 1.44 MB floppy
Microsoft Internet keyboard
Microsoft Intellimouse 3 button PS/2

The Maxtor HDs are both setup as Masters on individual IDE channels
of the Promise TX2 card.  The Samsung HD is on IDE channel 0 and
the NEC DVD drive is on IDE channel 1.

I see that your machine has two HighPoint HPT374 controllers and each
has 5 disks attached:

HPT374: IDE controller at PCI slot 03:01.0
HPT374: chipset revision 7
HPT374: not 100% native mode: will probe irqs later
HPT37X: using 33MHz PCI clock
    ide2: BM-DMA at 0x3000-0x3007, BIOS settings: hde:DMA, hdf:pio
    ide3: BM-DMA at 0x3008-0x300f, BIOS settings: hdg:DMA, hdh:pio
HPT37X: using 33MHz PCI clock
    ide4: BM-DMA at 0x3400-0x3407, BIOS settings: hdi:DMA, hdj:pio
    ide5: BM-DMA at 0x3408-0x340f, BIOS settings: hdk:DMA, hdl:DMA
HPT374: IDE controller at PCI slot 04:02.0
HPT374: chipset revision 7
HPT374: not 100% native mode: will probe irqs later
HPT37X: using 33MHz PCI clock
    ide6: BM-DMA at 0x4000-0x4007, BIOS settings: hdm:DMA, hdn:pio
    ide7: BM-DMA at 0x4008-0x400f, BIOS settings: hdo:DMA, hdp:pio
HPT37X: using 33MHz PCI clock
    ide8: BM-DMA at 0x4400-0x4407, BIOS settings: hdq:DMA, hdr:pio
    ide9: BM-DMA at 0x4408-0x440f, BIOS settings: hds:DMA, hdt:DMA


This is what I wanted to talk to you about.  You may be experiencing
degraded disk I/O performance by having one slave disk on each of the
HPT374 cards.

Since you told me that you were looking to use a different RAID card,
you may be willing to experiment with the current setup.  If you
are, we suggest removing each slave disk from the HPT374s, rebuilding
your RAIDs without them, and then run the same setup as you have
now (ingest and process all CRAFT data; ingest and file all NNEXRAD
data) to see if performance improves.

The other thing that caught my eye is the comment on the HighPoint
page about the HP374:

"Supports 32-bit 33MHz PCI bus"

I don't know if this means that it does not support 66Mhz PCI bus speeds.
Do you?  Notice that the Promise TX2 explicitly says that it supports
66Mhz PCI.

Another observation: we don't think that you gain much if anything
by formatting your RAID partition with EXT2 versus EXT3.  However,
we do know that specifying the same block size when formatting
a disk/RAID to the same size as the disk/RAID is set for _will_
improve efficiency.  This may be another thing to play with if
you decide to remove the 5th disk from your RAID controllers.

Cheers,

Tom
--
NOTE: All email exchanges with Unidata User Support are recorded in the
Unidata inquiry tracking system and then made publically available
through the web.  If you do not want to have your interactions made
available in this way, you must let us know in each email you send to us.

>From address@hidden  Wed Apr 28 13:20:21 2004

Howdy, Tom!

I believe the key points are the master/slave issue and the PCI 33 MHz 
bus speed.  The new car is ATA-100 fully capable, and has 12 ports. 
3Ware specifically says "one drive per channel" while Promist and High 
Point don't take that position.  In fact, I've documentation from both 
of the latter, telling you the sequence to plug up the slave drives.

I'm willing to play a little, but I am really desirous to get this thing 
into a no-playing mode.

I'll try to get it going tomorrow, as today's been decimated by 
administrivia and meetings, and I still have to load compilers and get 
MM5 and WRF running...

Regards,
gerry