Luke,
Thanks for the tips. I'm running FreeBSD 6.1 amd64, but, I can also
enable readahead on the raid controller, and also adaptive readahead.
Here's tests:
Readahead & writeback enabled:
bash-2.05b$ bonnie++ -d bonnie -s 16000:8k
Version 1.03 ------Sequential Output------ --Sequential Input-
--Random-
-Per Chr- --Block-- -Rewrite- -Per Chr- --Block--
--Seeks--
Machine Size K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP
/sec %CP
16000M 156512 98 247520 47 59560 27 83138 60 143588 21
792.8 7
------Sequential Create------ --------Random
Create--------
-Create-- --Read--- -Delete-- -Create-- --Read---
-Delete--
files /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP
/sec %CP
16 +++++ +++ +++++ +++ 27789 99 +++++ +++ +++++ +++
+++++ +++
,16000M,156512,98,247520,47,59560,27,83138,60,143588,21,792.8,7,16,+++++
,+++,+++++,+++,27789,99,+++++,+++,+++++,+++,+++++,+++
Writeback and Adaptive Readahead:
bash-2.05b$ bonnie++ -d bonnie -s 16000:8k
Version 1.03 ------Sequential Output------ --Sequential Input-
--Random-
-Per Chr- --Block-- -Rewrite- -Per Chr- --Block--
--Seeks--
Machine Size K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP
/sec %CP
16000M 155542 97 246910 47 60356 26 82798 60 143321 21 787.3 6
------Sequential Create------ --------Random
Create--------
-Create-- --Read--- -Delete-- -Create-- --Read---
-Delete--
files /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP
/sec %CP
16 6329 49 +++++ +++ +++++ +++ +++++ +++ +++++ +++
+++++ +++
,16000M,155542,97,246910,47,60356,26,82798,60,143321,21,787.3,6,16,6329,
49,+++++,+++,+++++,+++,+++++,+++,+++++,+++,+++++,+++
(As a side note- according to the controller docs, Adaptive read ahead
reads ahead sequentially if there are two reads from sequential sectors,
otherwise it doesn't).
So, I'm thinking that the RAID controller doesn't really help with this
too much- I'd think the OS could do a better job deciding when to read
ahead. So, I've set it back to no readahead and the next step is to look
at OS level file system tuning. Also, if I have time, I'll try doing
RAID 1 on the controller, and RAID 0 on the OS (or vice versa). Since I
have 6 disks, I could do a stripe of 3 mirrored pairs (raid 10) or a
mirror of two striped sets of 3 (0+1). I suppose theoretically speaking,
they should have the same performance characteristics, however I doubt
they will in practice.
Thanks,
Bucky
-----Original Message-----
From: Luke Lonergan [mailto:llonergan@xxxxxxxxxxxxx]
Sent: Wednesday, August 16, 2006 2:18 AM
To: Bucky Jordan; Vivek Khera; Pgsql-Performance ((E-mail))
Subject: Re: [PERFORM] Dell PowerEdge 2950 performance
Cool - seems like the posters caught that "auto memory pick" problem
before
you posted, but you got the 16GB/8k parts right.
Now we're looking at realistic numbers - 790 seeks/second, 244MB/s
sequential write, but only 144MB/s sequential reads, perhaps 60% of what
it
should be.
Seems like a pretty good performer in general - if it was Linux I'd play
with the max readahead in the I/O scheduler to improve the sequential
reads.
- Luke
On 8/15/06 1:21 PM, "Bucky Jordan" <bjordan@xxxxxxxxxx> wrote:
> Luke,
>
> For some reason it looks like bonnie is picking a 300M file.
>
>> bonnie++ -d bonnie
> Version 1.03 ------Sequential Output------ --Sequential Input-
> --Random-
> -Per Chr- --Block-- -Rewrite- -Per Chr- --Block--
> --Seeks--
> Machine Size K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP
> /sec %CP
> 300M 179028 99 265358 41 270175 57 167989 99 +++++ +++
> +++++ +++
> ------Sequential Create------ --------Random
> Create--------
> -Create-- --Read--- -Delete-- -Create-- --Read---
> -Delete--
> files /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP
> /sec %CP
> 16 +++++ +++ +++++ +++ +++++ +++ +++++ +++ +++++ +++
> +++++ +++
>
,300M,179028,99,265358,41,270175,57,167989,99,+++++,+++,+++++,+++,16,+++
> ++,+++,+++++,+++,+++++,+++,+++++,+++,+++++,+++,+++++,+++
>
> So here's results when I force it to use a 16GB file, which is twice
the
> amount of physical ram in the system:
>
>> bonnie++ -d bonnie -s 16000:8k
> Version 1.03 ------Sequential Output------ --Sequential Input-
> --Random-
> -Per Chr- --Block-- -Rewrite- -Per Chr- --Block--
> --Seeks--
> Machine Size K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP
> /sec %CP
> 16000M 158539 99 244430 50 58647 29 83252 61 144240 21
> 789.8 7
> ------Sequential Create------ --------Random
> Create--------
> -Create-- --Read--- -Delete-- -Create-- --Read---
> -Delete--
> files /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP
> /sec %CP
> 16 7203 54 +++++ +++ +++++ +++ 24555 42 +++++ +++
> +++++ +++
>
,16000M,158539,99,244430,50,58647,29,83252,61,144240,21,789.8,7,16,7203,
> 54,+++++,+++,+++++,+++,24555,42,+++++,+++,+++++,+++
>
> ... from Vivek...
> which is an issue with freebsd and bonnie++ since it doesn't know
> that freebsd can use large files natively (ie, no large file hacks
> necessary). the freebsd port of bonnie takes care of this, if you
> use that instead of compiling your own.
> ...
>
> Unfortunately I had to download and build by hand, since only bonnie++
> 1.9x is available in BSD 6.1 ports when I checked.
>
> One other question- would the following also be mostly a test of RAM?
I
> wouldn't think so since it should force it to sync to disk...
> time bash -c "(dd if=/dev/zero of=/data/bigfile count=125000 bs=8k &&
> sync)"
>
> Oh, and while I'm thinking about it, I believe Postgres uses 8k data
> pages correct? On the RAID, I'm using 128k stripes. I know there's
been
> posts on this before, but is there any way to tell postgres to use
this
> in an effective way?
>
> Thanks,
>
> Bucky
>
> -----Original Message-----
> From: pgsql-performance-owner@xxxxxxxxxxxxxx
> [mailto:pgsql-performance-owner@xxxxxxxxxxxxxx] On Behalf Of Vivek
Khera
> Sent: Tuesday, August 15, 2006 3:18 PM
> To: Pgsql-Performance ((E-mail))
> Subject: Re: [PERFORM] Dell PowerEdge 2950 performance
>
>
> On Aug 15, 2006, at 2:50 PM, Luke Lonergan wrote:
>
>> I don't know why I missed this the first time - you need to let
>> bonnie++
>> pick the file size - it needs to be 2x memory or the results you
>> get will
>> not be accurate.
>
> which is an issue with freebsd and bonnie++ since it doesn't know
> that freebsd can use large files natively (ie, no large file hacks
> necessary). the freebsd port of bonnie takes care of this, if you
> use that instead of compiling your own.
>
>
