Skip to contentSkip to conversation

2.5.63-mjb2 (scalability / NUMA patchset)

DISCUSSION [email protected] 2 participants 2 messages Mar 3, 2003

lore
Share & tools
lore
#0 lore raw
Quoted 4 lines

Pleeeeeeze remember to specify basic things such as the machine, amount of memory and especially the filesystem type in use.

Quoted 1 line

on uniprocessor K6/2 475 Mhz with 384 MB ram

and two IDE drives on ext2.

Could it be a disk driver issue? Maybe 2.4 has some IDE enhancements that aren't in 2.5 yet. Sounds crazy, but this is my reasoning...

On a quad Xeon with QLA2200, the QLogic driver makes the biggest difference for AIM7 dbase test.

Here is AIM7 dbase on quad Xeon with 3.75 GB ram over ext2:

AIM7 dbase workload

kernel                   Tasks   Jobs/Min         Real    CPU  
2.5.62-mm2               32     555.9            342.0   155.2 
2.4.21-pre4aa1           32     554.4            342.8   142.3 
2.4.21-pre4aa3           32     551.9            344.4   149.6 
2.4.21-pre4-ac3          32     473.8            401.2   147.7 
2.5.62                   32     473.6            401.3   148.2 
2.5.63-mjb2              32     472.5            402.3   161.5 
2.5.63                   32     471.6            403.1   153.1 
2.2.24-rc3               32     431.9            440.1   165.7 

2.5.62-mm2 has the feral driver. aa has the QLogic 6.x driver. Those two kernels rule AIM7 dbase and fserver on quad Xeon with QLA2200. I tested earlier 2.5 and aa with/without the newer QLogic drivers. It was _the_most_important_ factor for AIM7 dbase and fserver. Perhaps AIM7 dbase and fserver really suck. They seem rather impervious to other improvements in the kernel.

Quoted 1 line

Care to share your aim7 database methodology with me?

AIM7 dbase takes a mixture of AIM9 micro activities and runs them in proportion to what it's developers found a circa 1996 database running.

My methodology was:

0) Scratch head and look for benchmark that does more than one thing.
1) Stumble on AIM7 which purports to test database, fileserver, compute server and multiuser scalability.
2) Run AIM7 dbase to crossover
3) Find the lowest load where the Jobs/Minute is near it's peak.
4) Use result of 3) for max. Divide into 8 smaller loads.
5) Test all kernels with the same load.

Signature · 2 lines
Randy Hron
http://home.earthlink.net/~rwhron/kernel/bigbox.html
#1 lore raw
Quoted 6 lines · [email protected] wrote:
Quoted 5 lines

Pleeeeeeze remember to specify basic things such as the machine, amount of memory and especially the filesystem type in use.

Quoted 1 line

on uniprocessor K6/2 475 Mhz with 384 MB ram

and two IDE drives on ext2.

Ah, thanks.

Quoted 2 lines

Could it be a disk driver issue? Maybe 2.4 has some IDE enhancements that aren't in 2.5 yet.

Well I tested the AIM7 dbase workload yesterday on 256MB IDE. 2.4 and 2.5 have the same throughput, down to a fraction of one percent. The entire working set appeared to be around 200MB so there was no reading from disk at all. Just 25 minutes of trickling out very slow O_SYNC writes. The thing is dominated by disk seek time.

Quoted 13 lines

Here is AIM7 dbase on quad Xeon with 3.75 GB ram over ext2:

AIM7 dbase workload

kernel                   Tasks   Jobs/Min         Real    CPU  
2.5.62-mm2               32     555.9            342.0   155.2 
2.4.21-pre4aa1           32     554.4            342.8   142.3 
2.4.21-pre4aa3           32     551.9            344.4   149.6 
2.4.21-pre4-ac3          32     473.8            401.2   147.7 
2.5.62                   32     473.6            401.3   148.2 
2.5.63-mjb2              32     472.5            402.3   161.5 
2.5.63                   32     471.6            403.1   153.1 
2.2.24-rc3               32     431.9            440.1   165.7 
Quoted 1 line

2.5.62-mm2 has the feral driver. aa has the QLogic 6.x driver.

Well if there is any difference in drive caching policy then one would expect to see large differences. Using writeback caching in the disk (which is considered cheating) would speed things up. But I'd be surprised if 2.5-vs-2.4 IDE affected the drive's caching policy.

Quoted 5 lines

Those two kernels rule AIM7 dbase and fserver on quad Xeon with QLA2200. I tested earlier 2.5 and aa with/without the newer QLogic drivers. It was _the_most_important_ factor for AIM7 dbase and fserver. Perhaps AIM7 dbase and fserver really suck. They seem rather impervious to other improvements in the kernel.

Yes, they do.

Quoted 5 lines
Quoted 1 line

Care to share your aim7 database methodology with me?

AIM7 dbase takes a mixture of AIM9 micro activities and runs them in proportion to what it's developers found a circa 1996 database running.

AIM7 dbase would probably be more interesting if it created a larger working set.