Another Adaptec 2120s perfomance problem

[Scott Long]

kost wrote:
> Hi!
> 
> ---PROBLEM---
> I have got 2 servers running FBSD 5.x and RAID0.
> The first server has RAID 0 with 2 IDE drives and the second one has
> RAID0 with 8(!!!) SCSI drives.
> All test I have run on these platforms gave me almost same results. My
> testing was extremely simple and I expected to see at least 2 times
> faster perfomance on SCSI RAID... but it didn't happened.
> 
> ---HARDWARE---
> server1: P4 3Ghz, Intel 875, 1Gb RAM, Promise + RAID 0(2xIDE Seagate 80Gb). OS: FreeBSD 5.1
> 
> server2: P4 3,2Ghz, Intel 865, 1Gb RAM, Adaptec 2120S(7349 firmware) + RAID 0(8x SCSI
> Fujitsu 70Gb). OS FreeBSD(4.11, 5.2, 5.3, 5.4) and even Windows XP SP2
> Corporate.
> 
> ---TUNING---
> 1. Adaptec 2120s BIOS
> First time I have left almost all settings in default state but then
> I tried to turn off READ CACHE while creating ARRAY and turn on/off
> WRITE cache.
> 2. FreeBSD
> I'll show main FBSD settings for 5.3 RELEASE, because other versions had
> same performance results.
> 
> sqlmail# uname -a
> FreeBSD sqlmail.ropnet.ru 5.3-RELEASE FreeBSD 5.3-RELEASE #0:
> Fri Apr 22 17:00:01 UTC 2005     root at sqlmail.ropnet.ru:/usr/src/sys/i386/compile/SQL  i386
> 
> sqlmail# df
> Filesystem    1K-blocks   Used     Avail Capacity  Mounted on
> /dev/aacd0s1a    988398  51620    857708     6%    /
> devfs                 1      1         0   100%    /dev
> /dev/aacd0s1d   2971278 735948   1997628    27%    /usr
> /dev/aacd0s1f 550186622 260982 505910712     0%    /usr1
> /dev/aacd0s1e    988398    510    908818     0%    /var
> 
> sqlmail# kldstat
> Id Refs Address    Size     Name
>  1    3 0xc0400000 368e5c   kernel
>  2   14 0xc0769000 537f0    acpi.ko
> 
> Kernel was build in many variants: GENERIC, without aacp, without aacp
> and pass device and etc.
> 
> When aacp was enabled I got following strange results:
> sqlmail# dmesg
> pass0: <FUJITSU MAT3073NP 0105> Fixed unknown SCSI-3 device
> pass0: 160.000MB/s transfers (80.000MHz, offset 127, 16bit) 
> pass1 at aacp0 bus 0 target 2 lun 0 
> pass1: <FUJITSU MAT3073NP 0105> Fixed unknown SCSI-3 device 
> pass1: 160.000MB/s transfers (80.000MHz, offset 127, 16bit) 
> etc...

The aacp driver is just a simple SCSI passthrough transport that
allows one to access non-disk devices on the controller (like a
CDROM or tape drive) and flash drive firmware.  It is not needed
for normal operation, and in fact I discourage people from using
it unless they have a specific need.

> 
> Why 160MB/s if Adaptec bios and aacli show 320MB/s? Is it normal? And
> I saw nothing about "tagged queueing" enabled for at least one SCSI
> device... and then my suspicion confirmed:

This is a known bug in the driver.  If you want to find out the actual
speed of the disks, you'll need to use the aaccli tool.

> 
> sqlmail# negotiate 0:1 -v
> (pass0:aacp0:0:1:0): sync parameter: 8 
> (pass0:aacp0:0:1:0): frequency: 160.000MHz 
> (pass0:aacp0:0:1:0): offset: 127 
> (pass0:aacp0:0:1:0): bus width: 16 bits 
> (pass0:aacp0:0:1:0): disconnection is disabled 
> (pass0:aacp0:0:1:0): tagged queueing is disabled   (!!!!!) 
> aacp0: SIM/HBA version: 1 
> aacp0: supports 16 bit wide SCSI 
> aacp0: user has disabled initial BUS RESET or controller is in target/mixed mode    (!!!!!) 
> aacp0: HBA engine count: 0 
> aacp0: maximum target: 15 
> aacp0: maximum LUN: 8 
> aacp0: highest path ID in subsystem: 0 
> aacp0: initiator ID: 7 
> aacp0: SIM vendor: FreeBSD 
> aacp0: HBA vendor: Adaptec 
> aacp0: bus ID: 0 
> aacp0: base transfer speed: 3.300MB/sec 
> 
> I didn't disabled BUS RESET as like didn't set controller in target/mixed
> mode. I even don't now where to config these params, at least I
> couldn't find anything in Adaptec BIOS on this theme.
> 
> Of course all my attempts to turn on "tagged queueing" was failed.
> 

The passthrough interface doesn't support tagged queueing.  The RAID 
processor, however, is likely doing tagged queueing.  As for bus
resets, they are disabled in the aacp driver because they cause the
RAID processor to get upset.

The aacp device is just a simple hack for a specific purpose.  It has
a number of significant limitations, and is by no means a monitoring or
management tool for the AAC cards.  Please just remove it from your
configuration and forget that it exists.  The aaccli software is the
proper tool to use for management.

> 
> ---TESTING AND RESULTS---
> I understand that my tests are very far from ideal, but in my opinion
> I have to see better performance with SCSI RAID even in such test,
> especially we are talking not only about SCSI and IDE comparison but 2
> vs 8(!) spindles performance.
> Ok. what have I done:
> sqlmail# cd /usr
> sqlmail# fetch ftp://ftp.freebsd.org/pub/FreeBSD/ports/ports/ports.tar.gz
> sqlmail# tar -zxf ports.tar.gz
> sqlmail# time cp -pR ports.tar.gz /usr1/
> 0.577u 8.209s 2:12.44 6.6%      14+713k 41579+144io 75324pf+0w
> 
> While copying I run
> sqlmail# iostat 1 100
> on another console and got 3-5MB/sec speed for aacd0 device
> 
> sqlmail# systat -vmstat 1
> showed me 90-100% disk busy while cp was in progress
> 
> So... results were almost the same in all kernel variants, all FBSD
> versions and number of disks in SCSI RAID 0. And almost equal results
> I got on server1 with IDE RAID 0.
> 
> Also I tried to build RAID5+0... It was awful, because it takes twice
> more time to copy ports collection from one partition to another.
> 
> I tried to use another cable on 8 devices - same result
> ...tried to build ARRAY 0 with two SEAGATE CHEETAH - same result
> 
> Thank you for any answer.
> 

The 2120 is what Adaptec calls a 'value card'.  It is designed to be
very low-cost for those who want hardware raid.  Unfortunately, this
translates into it having pretty poor performance too.  On high-end
card, the driver and the OS performs quite well.

Scott