Tek-Tips is the largest IT community on the Internet today!

Members share and learn making Tek-Tips Forums the best source of peer-reviewed technical information on the Internet!

  • Congratulations IamaSherpa on being selected by the Tek-Tips community for having the most helpful posts in the forums last week. Way to Go!

Windows 2008 Disk Performance

Status
Not open for further replies.

technome

IS-IT--Management
Jun 9, 2003
1,860
US
Spent a few days with 2008 server Standard and Enterprise RC1 on a Dell 2900III, quad core, 8 Gig ram, Perc 6I raid adapter, setup as a raid 1 and raid 5 (4 disk), 15k SAS drives; aligned the partitions on both raids at 64. Disk benchmarked both Windows 2003 and 2008 on this server.....

Installation of 2008, including the necessary reboots took 20 minutes. Boot time <50 seconds after the bios screen. Win 2003 took about the same.

There is a BIG difference how both versions of Sever 2008 (and the Perc 6i) handle file caching compared to both Win 2000 and Windows 2003. The older server versions concentrated on large files throughput, generally the incorrect area to strive for, as most servers need to concentrate on small files. In a complete reversal, 2008 aims at small data chunks. The caching in 2008 is phenomenal, file sizes under 30 Meg are cached, disk benchmarking programs run from the cache, the disks rarely show access, unlike it's predecessors. Even IOmeter, with its build in access specifications only runs from cache without modifications.

The raid 1 performs very well, with the "write back" enabled, dismal with it disabled, as expected. The raid 5, "write back" enabled, performed better than the raid 1 in both reads and writes, writes are only slightly ahead, reads more so; I did not bother running raid 5 with "write through", a useless cause with any raid 5 setup.

I RCed to the desktop in admin mode and ran a few programs. Strangely, I kept looking up at the RDP screen tab, as it was hard to believe I was not at the console, no delays at all.

Played with the hyper-v....
As soon as this is installed, even if it is stopped, even if all the associated services are stopped... disk throughput drops approx 25%, due to the parameter changes in the registry. Installation of a VM running 2008 was a no brainer, but this is as far as I wanted to go, as this performance hit is beyond my acceptance, as this server will running among other things a disk intensive SQL based program. The good news, once hyper-v is uninstalled performance is back to pre-hyper-v levels

If I were to order the server now, 12 Gigs of ram would be my minimum. Did some registry tweaks, the usual for server 2003, on one install, which made no measurable difference.
The 32bit or 64 bit version, made no appreciable difference in disk throughput.


........................................
Chernobyl disaster..a must see pictorial
 
Do keep in mind that Windows 2008 is supose to align to 64 automatically, so we shouldn't need to diskpart the drives any more.

Denny
MCSA (2003) / MCDBA (SQL 2000)
MCTS (SQL 2005 / Microsoft Windows SharePoint Services 3.0: Configuration / Microsoft Office SharePoint Server 2007: Configuration)
MCITP Database Administrator (SQL 2005) / Database Developer (SQL 2005)

My Blog
 
I did not know that, I had aligned previously with 2003.

Found the alignment in server 2003 increased disk throughput in a 4 disk raid 5 with a 64k stripe, write back enabled, no read ahead about 5%
Raid 1, same parameters about 15%.




........................................
Chernobyl disaster..a must see pictorial
 
I've seen improvements with the adjusted alignment of up to 40% depending on disk load and the amount of IO being used. (High end SQL Servers and what not.)

Denny
MCSA (2003) / MCDBA (SQL 2000)
MCTS (SQL 2005 / Microsoft Windows SharePoint Services 3.0: Configuration / Microsoft Office SharePoint Server 2007: Configuration)
MCITP Database Administrator (SQL 2005) / Database Developer (SQL 2005)

My Blog
 
My tests were basically for disk throughput for varying files sizes and transfers between the arrays,I have yet to properly test for SQL speed. Windows 2008 (basically the default setup) is using 4 of the 8 meg of ram quite well for caching, unlike Win 2003 or 2000, which have major disk accesses on all file sizes during benchmarking. As I can see, the throughput numbers between 2008 and 2003 are similar; I am counting on 2008 caching ability to free up the raid adapter a good bit, especially for SQL. Now to get the client to get another 4 gig of ram.

Got a newsletter from Mark Minasi about changes in the TCP Window. I was wondering why 2008 downloaded Enterprise RC1 in 30 minutes, versus >3 hours on the next best machine on the network.


........................................
Chernobyl disaster..a must see pictorial
 
I take it back about IOmeter, just did a fresh install so I could document all setting before any modification, it does access the disk.

........................................
Chernobyl disaster..a must see pictorial
 
Wow,
I took a look at the chernobyl pictures. It is shocking.

Dean

Dean

forum.gif

Online Screencasts and Video-Tutorials
 
Yes shocking, and I am sure this type of accident will happen again. Wonder if Russia will ever release the true death toll.



........................................
Chernobyl disaster..a must see pictorial
 
I finally got tired for IOmeter's odd behaviors, so I tested disk IO with SQLIO....
In disk IOs using SQLIO, raid 1 produced approx 2-3% higher IOs overall, then the raid 5 due to a couple data chunk sizes; most data sizes produced equal benchmark results due to caching. SQLIO parameters were at 8 outstanding IOs, both random and sequential reads and writes were tested at sizes 256,128,64,32,16,8,4,2,1 .

Hyper-v...
Contrary to previous testing results, disk throughput and IO tests results were very close to pre-Hyper-v levels (without any VMs installed). With the last testing, I downloaded the newest bios release, and did not attach the network interface, did not have the time to test further. I believe the Broadcom driver had affected the previous testing.
2008 appears to leave no registry parameter changes in place from role or feature installations once they are removed; if only third party software adhered to this !!!!

Array Stripe size....
Tested disk throughput and IO with a 64k and 128k array stripes, most of the disk/IO benchmarks from different benchmark programs show variations according to data chunk size; overall it appears the stripe size changes result in a small difference in overall disk throughput/IO, the edge going to the 128k stripe for data chunk sizes as dictated by SQLIO.exe, and disk throughput tests. What really complicates the benchmarking, many of the tests are actually running in cache or are highly affected by the raid adapter/OS combination.; certain data chunks perform higher on a 128k stripe than a 64k stripe and other sizes in the reverse. I will likely leave it on the 64k stripe once this is setup for production, as this server will be used for databases and as a file server; as with any stripe size change there are always trade offs. I would like to have tried a 32k size, but the benchmarking is very time consuming/beyond tedious. As is, with Windows caching doing a nice job on small data chunks for any raid type, I can not see any reason for going to a raid stripe other than the default,in my situation. If this were a dedicated database or Exchange server, I would retest.

With so many changeable parameters in 2008, I print screened all the default setting throughout 2008. With so little info out there, I know I will need to reference the settings (after I run into trouble). There are just too many not to have a document on hand.


........................................
Chernobyl disaster..a must see pictorial
 
Paul,

Would you like to post the results in an article on NetoMeter? I will start publishing (besides the screencasts)short articles with benchmarks and opinions with a presentation of the author in the beginning.

Dean

Dean

forum.gif

Online Screencasts and Video-Tutorials
 
Will get back to you in a few days, I just got called in on a network which was massively hacked, nothing like 18 hour days.

Thanks
Paul Meiners

........................................
Chernobyl disaster..a must see pictorial
 
Dean
Will pass on the article, I am so far behind in work load due to the hacked network; have to say the hacker provided me with some intesting work, more like a chess game...I took over this network after the hacker gained access, guy was good, but in the end he lost.
Thanks anyway

........................................
Chernobyl disaster..a must see pictorial
 
No Problem, Technohome.

I am glad you were able to solve the problem.

Dean

Dean

forum.gif

Online Screencasts and Video-Tutorials
 
Status
Not open for further replies.

Part and Inventory Search

Sponsor

Back
Top