DBT3-pgsql large performance improvement 2.6.6-rc1
Sorry, forgot to copy the list.
-----Forwarded Message-----
From: Mary Edie Meredith <[email protected]>
To: Andrew Morton <[email protected]>
Subject: Re: DBT3-pgsql large performance improvement 2.6.6-rc1
Date: Mon, 19 Apr 2004 08:44:45 -0700
The full reports for the 8ways are at:
2.6.5 http://khack.osdl.org/stp/291346/
2.6.6-rc1 http://khack.osdl.org/stp/291915/
Most stats are broken up between the load phase, power (single stream), and throughput phases. By the way the load phase improved from 1:02:43 (HH:MM:SS) to 0:41:41 as well.
Another data point that I noticed is that a similar boost in performance occurred at 2.6.5-mm5 compared to 2.6.5-mm3 (there were problems with -mm4 so no data exists). The 8way results for that are at:
http://khack.osdl.org/stp/291727/
Would the radix tree changes only apply to writing? Don't the searches for writable pages occur whether there are pages to write or not? There is a considerable amount of memory devoted to page cache in this set up.
Quoted 59 lines · Andrew Morton
Quoted 14 lines · Mary Edie Meredith
Performance in DBT-3 (using PostgreSQL) has vastly improved for
_bothin the "power" portion (single process/query) and in the "throughput" portion of the test (when the test is running multiple processes) on our 4-way(4GB) and 8-way(8GB) STP systems as compared 2.6.5 kernel results.Using the default DBT-3 options (ie using LVM, ext2, PostgreSQL version 7.4.1)
Note: Bigger numbers are better.
Kernel....Runid..CPUs.Power..%incP.Thruput %incT
Quoted 8 lines
2.6.5 291308 4 97.08 base 120.46 base 2.6.6-rc1 291876 4 146.11 50.5% 222.94 85.1%Kernel....Runid..CPUs.Power..%incP..Thruput %incT
2.6.5 291346 8 101.08 base 138.95 base 2.6.6-rc1 291915 8 151.69 50.1% 273.69 97.0%So the improvement is between 50% and 97%!
How odd.
Quoted 4 lines
Profile 2.6.5 8way throughput phase:
http://khack.osdl.org/stp/291346/profile/after_throughput_test_1-tick.sort
Profile 2.6.6-r1 8way throughput phase:
http://khack.osdl.org/stp/291915/profile/after_throughput_test_1-tick.sortOdder.
do_anonymous_page()is doing 10x more work in 2.6.6-rc1. And the CPU scheduler cost has fallen a lot.Frankly, I can't think of anything in 2.6.6-rc1 which would cause either of these things!
Quoted 4 lines
What I notice is that
radix_tree_lookupis in the top 20 in the 2.6.5 profile, but not in 2.6.6-rc1. Could theradix tree changes be responsible for this?I would certainly expect 2x or even higher throughput increases from either the writeback changes or the ext2&ext3 fsync changes.
Quoted 8 lines
DBT-3 is a read mostly DSS workload and the throughput phase is where we run multiple query streams (as many as we have CPUs). In this workload, the database is stored on a file system, but it is small relative to the amount of memory (4GB and 8GB). It almost
completely caches in page cache early on. So there
is some physical IO in the first few minutes, but very little to none in the remainder.But you're not doing a significant amount of writing during the test, so scrub that theory.
Do you have full reports anywhere? I'd be interested in seeing a vmstat trace from the entire run, both 2.6.5 and 2.6.6-rc1.
Signature · 4 lines
Mary Edie Meredith [email protected] 503-626-2455 x42 Open Source Development Labs
Mary Edie Meredith wrote:
.snip
Quoted 12 lines
Quoted 12 lines
Quoted 9 lines · Mary Edie Meredith
Performance in DBT-3 (using PostgreSQL) has vastly
improved for_bothin the "power" portion (single process/query) and in the "throughput" portion of the test (when the test is running multiple processes) on our 4-way(4GB) and 8-way(8GB) STP systems as compared 2.6.5 kernel results.Using the default DBT-3 options (ie using LVM, ext2, PostgreSQL version 7.4.1)
Note: Bigger numbers are better.
Kernel....Runid..CPUs.Power..%incP.Thruput %incT 2.6.5 291308 4 97.08 base 120.46 base 2.6.6-rc1 291876 4 146.11 50.5% 222.94 85.1%
Quoted 7 lines
Quoted 7 lines
Quoted 5 lines
Kernel....Runid..CPUs.Power..%incP..Thruput %incT
2.6.5 291346 8 101.08 base 138.95 base 2.6.6-rc1 291915 8 151.69 50.1% 273.69 97.0%So the improvement is between 50% and 97%!
How odd.
Your instincts were right (as usual).
We took another closer look at this. It appears that a change slipped into the kit around April 13, which probably made this stunning performance difference not the kernel.
Results of second run of the linux-2.6.5 kernel using the changed test kit showed that the numbers are comparable to the latest 2.6.6-rc3 release. Therefore,
it looks like the initial improvements posted were most likely due to the change in datatypes making the database size smaller, among other things.
....................Composite....Power...Throughput
initial linux 2.6.5 108.14 97.08 120.46 second linux 2.6.5 180.62 146.53 222.63 linux 2.6.6-rc3 181.83 147.16 224.68
Tests were compared on 4-way systems. (Bigger is better).
We will run the 8-way numbers over night to see if there is any difference for those systems. We are also running some IA64 results over night as well.
Quoted 27 lines
Quoted 27 lines
Quoted 6 lines
Profile 2.6.5 8way throughput phase:
http://khack.osdl.org/stp/291346/profile/after_throughput_test_1-tick.sort
Profile 2.6.6-r1 8way throughput phase:
http://khack.osdl.org/stp/291915/profile/after_throughput_test_1-tick.sortOdder.
do_anonymous_page()is doing 10x more work in 2.6.6-rc1. And the CPU scheduler cost has fallen a lot.Frankly, I can't think of anything in 2.6.6-rc1 which would cause either of these things!
Quoted 1 line
What I notice is that
radix_tree_lookupis in the top 20 in the 2.6.5 profile, but not in 2.6.6-rc1. Could theradix tree changes be responsible for this?I would certainly expect 2x or even higher throughput increases from either the writeback changes or the ext2&ext3 fsync changes.
Quoted 1 line
DBT-3 is a read mostly DSS workload and the throughput phase is where we run multiple query streams (as many as we have CPUs). In this workload, the database is stored on a file system, but it is small relative to the amount of memory (4GB and 8GB). It almost completely caches in page cache early on. So there is some physical IO in the first few minutes, but very little to none in the remainder.
But you're not doing a significant amount of writing during the test, so scrub that theory.
Do you have full reports anywhere? I'd be interested in seeing a vmstat trace from the entire run, both 2.6.5 and 2.6.6-rc1.