[THIN] Re: RAID 1

Andrew:

You're the best!  Thanks for listening and weighing my contribution to the 
string -albeit I should have used the designation Mirror and Stripe rather than 
null and one; thankfully you understood my message [hopefully I was consistent 
in err ;) ]  Thankfully we have people and processes whose algorithm includes 
perception...  

 

NUTSHELL:  Striping performance over mirror is NOTICABLE.  Try this write up:  
http://www.microsoft.com/technet/prodtechnol/windows2000serv/reskit/core/fndb_con_dwtx.mspx?mfr=true
  I guess, the most applicable documentation for you would be to image your 
server as is, then setup your disks mirrored and then drop down the image and 
setup a performance monitor.  Next you setup your disks striped and setup the 
same performance monitor.  Finally for a baseline setup the server on one 
single disk (cut up the disk to emulate the patricians) and record with 
performance monitor -that is if you are truly interested in knowing actual 
performance gains applicable to your environment as the only change (remember 
you are using an image) is the disk configuration.  You asked, there's the real 
answer -no formula can tell you exactly how your server is going to perform 
until you slap down all the applications and configurations on the same box. 

 

You probably have seen the wiki 
(http://en.wikipedia.org/wiki/Standard_RAID_levels), there are some formulas 
for you to scribble down and put in your ROI documentation.  I suppose factors 
that would influence the write up and is not documented would be performance 
gain from disk speed and caching.  

 

But from personal experience supporting 1000-10,000 employee companies and 
enterprise internet forward-facing NT clustered environments (we average being 
in the top 100-500 on the internet) [that is NOT me boasting but just 
information about my background and environments I have been in -the solutions 
and configurations I have tried successfully- I put it in there to give you an 
idea of what type of situations I have built and supported, I do not know where 
you are coming from and what types of constraints you have] Striped NOS volumes 
with a dedicated disk for temp files have consistently and noticeably out 
performed data-redundant volumes [a separate patrician will not be as fast for 
temp files but yes, it will give you a performance nudge.]

 

What can you do though about a disconnected Citrix user -no matter what they 
are working on they will be miffed.  What happens to fifty disconnected users 
is Citrix issue at heart not necessarily a Hardware, Software or configuration 
change issue; elaborated:  Say if you have an environmental variable beyond 
your control at location A, Citrix is not intelligent enough to seamlessly 
transfer a User and their Environment to location B or the Citrix Technology 
may not have the opportunity to transfer the User and Session to another 
location in time; therefore you will ALWAYS have the possibility for downtime.  
 Every computer (regardless of workstation or server) can crash or have the 
unforeseen occur to it -you need to be able to convey that to your end users 
and clients.  NOTHING ever in computers is absolute.

 

I understand why Citrix would be the dominate topic; however, I thought this 
was a Terminal Server string -so that's why I offered the configuration. 

                

Joe:

You will have to read the documentation that came with your box to configure, 
but here is the concept to your answer:

http://www.microsoft.com/technet/prodtechnol/windows2000serv/reskit/core/fndb_con_dwtx.mspx?mfr=true

Table   11.3 Structure of a Striped Volume

Stripe Number

Disk 1

Disk 2

Disk 3

Disk 4

Stripe 1

1 

2

3

4

Stripe 2

5

6

7

8

Stripe 3

9 

10

11

12

Stripe 4

13 

14

15

16

Stripe 5

17 

18

19

20

                                        

 

Some places are setup for PXE boot (meaning they boot off a SAN so no disks are 
needed on the server itself); therefore the disks are not necessary on the 
blade server and you can dedicate as many disks and configure the disks in 
whatever manner you choose.

 

 

Mike Dupris

 

From: thin-bounce@xxxxxxxxxxxxx [mailto:thin-bounce@xxxxxxxxxxxxx] On Behalf Of 
Andrew Wood
Sent: Tuesday, May 01, 2007 3:47 AM
To: thin@xxxxxxxxxxxxx
Subject: [THIN] Re: RAID 1

 

Well - I knew you meant RAID0 when you wrote RAID1 and vice versa - I see what 
you did there, keep everyone on their toes and make sure we're paying 
attention. I see I'm not the first to comment, so I take it I've not won the £5 
L

 

My "issue" is that for blades/1U servers (which is not an uncommon config for a 
citrix/ts server) what *is* the performance difference for a 
RAID1/RAID0/separate drives - is it significant? 

 

Yes, technically, when you've multiple spindles you get a better performance - 
but, what does that better performance relate to in the real world? You've only 
two drives - how much 'better' is 'better'? 

 

I would typically configured a RAID1 config as it's the more reliable - but, 
what I've not had chance to do is take a couple of servers, build 
raid1/raid0/separate drive configs and then load them up which performs better. 

 

And then

 

Transactions or archives should not be kept on Raid 1.  RAID 1 should not be 
used if the application's "presence" is not managed for you (Clustered VIP, 
Farm) -I guess I am saying, if the box goes down and you have another host that 
is going to handle the load then you shouldn't be afraid to use RAID 1 over 0.  

 

Yes - that's right and dandy for, say, a SQL or exchange backend - but perhaps 
that's of little use to you if your citrix server goes off-line with 50 
concurrent users. Punters may well be able to reattach on another server in the 
farm - but only after they've hunted you down and maimed you with a rusty spoon 
as they've lost the end of year financial data they were working on at the 
time... yet ...is the risk of that happening worth taking given the (possible) 
performance increase that you've gained by dropping raid1 and using raid0?

 

I'm more than willing to accept a short contract for me to muck about with this 
and bring back findings if anyones interested J

 

From: thin-bounce@xxxxxxxxxxxxx [mailto:thin-bounce@xxxxxxxxxxxxx] On Behalf Of 
Dupris, Mike
Sent: 30 April 2007 22:30
To: thin@xxxxxxxxxxxxx
Subject: [THIN] Re: RAID 1

 

Raid 1 out performs Raid 0 but is riskier because of the lack of redundancy. 
Raid 1 is better best used for basic "stuff" that you can easily get back on 
due to failure (NOS, easy applications quickly installed) -think if you can 
rebuild it from disk (or blast down from an image) then do a basic restore of a 
limited amount of settings or configurations (from file) and your operation is 
back up and running then RAID 1 is better for performance.

 

Raid 0 is delayed because of the read & compare then delete & writes.  Raid 1 
spreads data across disks thereby spreading out the disk operations across the 
array (relieving bottlenecks.)

 

Temporary files and system pagefiles can be kept on a single independent disk 
(rather than the same disk -remember we are trying to limit the times we have 
to wait on the arm to position the head on one of the platters -for the matter, 
we are also wanting to remove overhead on an array controller.)  Risk:  if that 
disk the temp/system file is located on fails during production, that could 
cause a serious issue if your system relies upon heavy pagefile use. 

 

Transactions or archives should not be kept on Raid 1.  RAID 1 should not be 
used if the application's "presence" is not managed for you (Clustered VIP, 
Farm) -I guess I am saying, if the box goes down and you have another host that 
is going to handle the load then you shouldn't be afraid to use RAID 1 over 0.  

 

Here is a high performance production setup for Clustered or Farmed boxes

 

[NOS] 3 disks, raid 1, controller A, disk cache enabled, PCI bus X (I like at 
least 3 for dedicated OS ,volume -single patrician)

[TEMP] 1 disk, independent disk, disk cache enabled, controller B, PCI bus Y

[DATA] Multiple disks, raid 5 or 10, disk write-through if Database, controller 
C or SAN attached, PCI bus Z or SAN attached

 

Where XYZ are placeholders for whatever dedicated PCI buses are available on 
your system. 

 

Mike Dupris

 

From: thin-bounce@xxxxxxxxxxxxx [mailto:thin-bounce@xxxxxxxxxxxxx] On Behalf Of 
Andrew Wood
Sent: Monday, April 30, 2007 7:45 AM
To: thin@xxxxxxxxxxxxx
Subject: [THIN] Re: RAID 1

 

I've always wanted to do this but never had the time - it'd be interesting to 
note what the performance gain was in relation to the fact that you increase 
your risk of causing an outage due to a single disk failure...

 

From: thin-bounce@xxxxxxxxxxxxx [mailto:thin-bounce@xxxxxxxxxxxxx] On Behalf Of 
Joe Shonk
Sent: 28 April 2007 15:38
To: thin@xxxxxxxxxxxxx
Subject: [THIN] Re: RAID 1

 

It's a very interesting question indeed.   Which blades are you using?  What 
kind of i/o controller?

Putting the pagefile (alone) on a second partition will help performance with 
regards that the pagefile will be create as one large continuous(non 
fragmented) file. 

With some controllers (those with NO cache), you can see an increase in 
performance by not using RAID 1 and have two independent spindles and put the 
pagefile, temp dirs, and spooler on the second drive.  But if you are going to 
do that, then why not set it up a RAID 0 with a second partition for the 
pagefile.  We literally 2x the performance in reads and writes with the 
configuration over RAID 1. 

Joe

On 4/27/07, Charles Watts <gregwatts77@xxxxxxxxxxx> wrote:


We are using HP Blades with RIAD 1 for our Citrix Servers on Windows Server
2003 SP1. My question is would we get better disk i/o performance if we
partioned the RAID 1 into two partitions and stuck the page file, temp 
directories and programs on the second partition? If not has any one used a
RAMDISK for this? Or is their an argument for eliminating RAID 1 altogether
and just put the page file, temp directories on the second hard drive 
(better performance and more disk space). Since I have a boat load of blades
where is the risk? So I lose one or two drives a year on my servers and each
time 50 - 60 users get kicked off. heck! I do that accidently at least once 
a year!?  Managment might not buy that argument but If the performance gain
is significant then it's worth it. Your thoughts?

Thanks,

Greg


SBC SITES ONLY GOOGLE SEARCH: http://www.F1U.com
************************************************
For Archives, RSS, to Unsubscribe, Subscribe or
set Digest or Vacation mode use the below link:
http://www.freelists.org/list/thin
************************************************

Other related posts: