Sign in with
Sign up | Sign in
Your question

AMD Vs Intel 4 Specific Server Apps.

Last response: in CPUs
Share
February 6, 2006 8:54:02 PM

Hey People. I wasn't sure where to post this discussion, either on the networking side or hardware....

But anyway.
I work for a small local government council. We have about 70 users total.
30 - 40 Terminal server.
40 - 30 Local Network.

Currently we have only ONE server. Which handles everything, DNS, DHCP, Finance Apps (Interbase), Records Management (SQL), Print Server, File Server, Mail Server, and the real hardware killer is the Terminal Server (25 users average. 40 users Peak)

The server we are using is under HUGE load as you could probably guess. Its a Dual CPU Intel Xeon 3Ghz. 2gig Ram. 4 SCSI Drive (not in RAID Config)

Obviously we are looking at upgrading. And I guess the main question is, when it comes to situations like this, what is the best CPU/s for the Job?

AMD Vs. Intel.

Could you please advise! (Please don't turn this into a INTEL Vs. AMD fight)

Just would like to know the facts!

Peace.[/b]
February 6, 2006 9:00:23 PM

For server applications, I would have to go with AMD. They generally have better throughput, are a bit quicker with multiple instructions(shorter paths) in the server side of things and they generally multitask well.

Most of it depends on the boards you use, the chipsets, quality of server hardware etc. AMD would probably be your best bet. I am one that rides on the fence using both processor brands but based on experience in the server market, AMD seems to shine.
February 6, 2006 9:09:13 PM

Quote:
Upgrade?

You guys needs 3 servers.
Keep the Xeon you guys already have got now and build another dual Opteron. Gather some old P3 scrap computer for the dedicated DHCP server.


I aggree with you totally!!!!
I am the only IT guy here, i'm only 19 and have been here for under a year. I have been telling them they need 3 servers to split the load for a very long time now.
But because of my age they wont listen to me!!
We are keeping the Xeon and using the new server for the terminal server. Possibly put a NAS box for the file storage as well.

Its a very political place, it consists of 2 rival towns with a massive VPN over the two. Very hard to get anything done.
Related resources
February 6, 2006 9:18:59 PM

Quick question:

Do you plan on actually building a new server? If so- Opteron is the way to go.

If not, you're probably going to find Xeon servers for a cheaper price. Price performance wise, Opteron has the Xeon beat. But since Opteron servers aren't that well marketed or priced by the big boys- you might be forced to go w/ a new Xeon server.

Like Wusy and other said though- you'd be better off with seperate servers for each task. Not only does it build redundancy it also will improve performance across the entire network.
February 6, 2006 10:08:01 PM

:D 

I would say either AMD64 939 or AMD64 940 or a combination thereof.

For tasks that require a lot of CPU and a lot of IO I would use a Dual Opteron. For tasks that are not very CPU and IO intensive I would use an AMD64 939 + nForce4 939.

For a Dual Opteron I would use something like this as a base config:

http://secure.newegg.com/NewVersion/Wishlist/WishShareS...

The S2882 is a very special board, not only do the Opterons each have 4 dedicated DIMM slots which can both run Dual Channel, they also have 2 independent PCI-X buses which greatly increase the IO bandwidth which is perfect for fast storage with RAID5 or 6 controllers or other high-end storage.

A 3Ware 9550SX SATA II CTRL can offer 800MB/sec reads and 380MB/sec writes in RAID5 :D 

For a single 939 something like this:

http://secure.newegg.com/NewVersion/Wishlist/WishShareS... minus the 7800GT of course ;-)

I would run Linux or *BSD on all of them because both are 100% legally free, open source, fast, easy to setup and maintain and relatively very secure.
[code:1:4fa489afd6]
Maybe use:

0. one nice 939+nForce4 for email + anti-virus + anti-spam - Linux

1. one nice 939+nForce4 or 940 for SQL and maybe Finance Apps - Linux

2. one nice 939+nForce4 or 940 for File and Print serving using Samba, etc - Linux

3. one low-end 939+nForce4 for DNS, DHCP and TFTP Linux or *BSD.
[/code:1:4fa489afd6]

Newegg would be the best source for the parts!

I have newegged a lot of server parts and built many servers myself for less than HALF of what Dell wanted for a similar sever! :D 

Semper Fi Linux on!

PS Check these disk benchmarks on a Dual 252 + S2882 + 3Ware RAID5 :D  https://66.235.243.163/bench/index.html [FC3 i386 LEFT] [FC3 x86_64 RIGHT] both of the same hardware.

PPS Linux or BSD will save you 5K or MORE on software licensing fees alone!
February 7, 2006 12:45:58 AM

Thanks for all your help Guys!
I wouldn't mind building the server myself, considering my background is hardware (not this networking arrrghhh)

I'm in Australia so we don't have NewEgg here, sadly.

As for Linux, I do like Linux, but my background as well as the external contractors background does not include Linux.
Yes it would save heaps on Licensing, however its not my money, so it doesn’t really worry me :wink:

I'm not sure who to go through, my main concern is redundancy, IBM seem very reliable from my experience.
What do you guys think?

I also wish to go Opteron, however the partners I'm dealing with view AMD as a poor unreliable company (VERY hard to pursued).
February 8, 2006 6:02:12 AM

Quote:
Thanks for all your help Guys!
I wouldn't mind building the server myself, considering my background is hardware (not this networking arrrghhh)

I'm in Australia so we don't have NewEgg here, sadly.

As for Linux, I do like Linux, but my background as well as the external contractors background does not include Linux.
Yes it would save heaps on Licensing, however its not my money, so it doesn’t really worry me :wink:




I would strongly advise you to consider Linux and or BSD. Open source has many benefits, so many in fact that large companies like IBM and HP are encouraging users to switch to Linux and are both making products which run Linux.

I do like the fact both of them support Linux however I would recommend custom built machines if possible for best results.


Quote:

I'm not sure who to go through, my main concern is redundancy, IBM seem very reliable from my experience.
What do you guys think?

I also wish to go Opteron, however the partners I'm dealing with view AMD as a poor unreliable company (VERY hard to pursued).


This is absolutely 10000000% NOT true. AMD is just as reliable as Intel if not MORE so.

I have built many AMD desktops and servers over the last few years and none of my AMDs had any serious problems. I know several major ISPs that use AMD exclusively and they have machines that have been up and running for YEARS without any problems whatsoever (particularly the ones running Linux and BSD). http://serverbeach.com/catalog/ for example was 100% AMD until they changed ownership recently and added some Dells for marketing reasons.

Dell itself is considering building AMD machines because they keep losing out to the competition.

It is not unusual at all to see Unix machines have uptimes of 1, 2, 3 or more years (granted you should be patching your kernel - which requires a reboot but if you do not mess with your kernel and just patch the rest of the system it will keep going for ever like the Energizer bunny).

In THG's endurance tests the Intel machine kept crashing while the AMD machine kept going.


I could build you just about any server you want however getting it down under might be a bit challenging ;-)

http://penguincomputing.com makes some great machines but I'm not sure if they ship to .AU If you cannot have your machines custom built ( which would be ideal ) I would advise you to look into ordering from them.
February 8, 2006 6:29:07 AM

Out of the 500 top supercomputing clusters in the world at LEAST 45 of them use AMD Opteron CPUs.

http://www.top500.org/lists/2005/11/basic


In fact the Los Alamos National Laboratory has over 2816 Opteron machines (those are the ones they are willing to acknowledge publicly. If Uncle Sam is willing to trust some of our most crucial research to AMD machines I think you can too :D  :D 

Most of these supercomputers run Linux, BSD or Unix btw
February 8, 2006 7:24:11 AM

Hello all!

Interesting topic,i just want to add my 2 cents to the bulk.
Well, opterons currently slightly beat the xeons ,but it's still "slightly" so you don't loose much by going xeons.
And the statement "AMD is unreliable" is quite understandable since Intel have usually dominated this area(untill now).
The stress test is irrelevant BTW.
There is a big variaty of Xeons from cheap to expensive as hell.
But what i'm trying to underline here is that opterons are not much better. Their main benefit is the low consumption, which counts a lot,but can be left aside(considering the lean towards Intel).
Every day users LOVE the opterons since they clock them to FX speed, making the ultimate gaming machine. But since we're talking servers and probably no overclocking both would be suitable to the task, though one will be consuming more energy than the other.
Price wise,i think they are close valued with 50-200 $ difference. But that's about the price you pay for going Intel instead of AMD(and for having an Intel fan as boss) 8)
P.S.
I agree that Opteron >Xeon but it is still hard to believe that duel video on the AMD site since that was created by AMD guys. AMD can play dirty too, you know.
February 8, 2006 7:46:27 AM

There are plenty of benchmarks which prove the Opteron is better and faster in many apps.

Here are some recent results:

http://www.spec.org/cpu2000/results/res2006q1/

Compare these results:

[code:1:506ed25185]

Acer Incorporated Acer Altos R710 (Dual-Core Intel Xeon 2.8 GHz, 2x2MB L2, 800 4 cores, 2 chips, 2 cores/chip(Hyper-Threading Technology Enabled) 58.6 58.6

Fujitsu Siemens Computers PRIMERGY BX630, AMD Opteron (TM) 875 4 cores, 2 chips, 2 cores/chip 63.9 69.8

[/code:1:506ed25185]

and these:

[code:1:506ed25185]

Dell Precision Workstation 380 (3.8 GHz Pentium 4, 2MB L2) 1 core, 1 chip, 1 core/chip (Hyper-Threading Technology disabled) 2088 2091

Fujitsu Siemens Computers CELSIUS V830, Opteron (TM) 254, Linux 64-bit 1 core, 1 chip, 1 core/chip 2094 2306

[/code:1:506ed25185]

and specviewperf

http://www.spec.org/gpc/opc.data/vp81/summary.html


There are also real-world benchmarks which prove the Opterons run very well particularly under Linux x86_64.

Here are some disk benchmarks I ran myself on a Dual Opteron 252 with HW RAID5

FC3 i386 is on the LEFT

FC3 x86_64 is on the RIGHT

https://66.235.243.163/bench/index.html

In 64bit mode the Opterons perform a LOT better.

The same is true even under XP64 and 2003 64 even though Linux is way better/faster/cheaper.
February 8, 2006 8:03:50 AM

Agreed,agreed agreed.
but the fact that the product is an "Intel" is enough to prove its cost(perf/watt and price).
Or is it?
Well depends on the person who buys it.
AMD still has quite a bad reputation IMO, and i tell you from experience.
My parents are both engineers and were they work AMD is forbiden,no matter the application. All schools in my town have labs filled with Intel procs.
Intel while bringing expensive performance doesn't let you down(the last reason for a malfunction would be the Intel proc or mobo ). AMD does it too but only for a couple of years.
Well,i have to admit i like Intel myself(i even have a 12 year old laptop( with Intel proc.) running and all my Intel based rigs have never let me down) but believe me,that's the case. Reputation is a thing that you can't destroy in a year(2005 was probably the worst year for Intel) and companies prefer not to "risk" so they go Intel. It takes years for that to change. If Conroe/Merom will do what they are supposed,2005 will surely be forgotten and forgiven.
February 8, 2006 8:47:22 AM

My AMD 386DX-40 still runs fine after about 16 years or so.

I call that reliability!

I own a lot of Intel CPUs however I prefer AMD hands down.

My 2 primary systems are AMD64s. My 2 backup systems are AMD32s. The rest are AMD32, Geode, Intel PIII, Intel PII, Intel Pentium Pro, Intel Pentium, i486DX-33, i386SX-25, AMD386DX-40, etc
February 8, 2006 9:25:40 AM

Lets look at the apps and the real issues, CPU choice should not be your real concern.

Quote:
Currently we have only ONE server. Which handles everything, DNS, DHCP, Finance Apps (Interbase), Records Management (SQL), Print Server, File Server, Mail Server, and the real hardware killer is the Terminal Server (25 users average. 40 users Peak)


Firstly you are at real risk only having one server, if that goes down all your apps are down (what are you using to backup the systems?)

Your first priority needs to be data security, consider buying a second system of roughly the same spec and cluster the two (MS cluster) or go for a cluster in a box (HP do a good one based on a DL560).
One half of the applications could run on one half of the cluster the other on the other half and fail over accordingly.

Personally I would do the following.
Look at a NAS filer (Network attached storage), for all data storage and some form of resiliance
A couple of small boxes for DNS and DHCP and domain controllers. (do not need top be clustered as backup IP addresses are configured anyway)
Then print/file/mail and at a push database based apps on a clustered pair.
Terminal server is a real killer and needs lots of memory. So have this on its own.

Now you need to ask the difficult questions, to get a decent budget.
How much does it cost if the system is down for
a) an hour
b) a day
c) a week

Then how much data loss can you afford
a) and hours worth
b) 12 hours worth
etc, etc.


The answeres will determin your budget and should be entertaining watching them squirm when they realise under how much risk they are with the current setup.
February 8, 2006 9:29:19 AM

Hey Linux, I see you been spreadin that Penguin Disease *wink wink*.

For the topic: of the Xeon vs. Opteron experience I've had, I've been around 7 Xeon Systems and 2 Opteron Systems, all Dual CPU. The Opteron Performance/Bandwidth has always been far superior to that of the Xeon, mainly because of the FSB vs. HTT for CPU Coherency (1 CPU talking to another) and AMD64's OnBoard Memory Controller. I wouldn't go as far as to say Xeon should never be used (some people around me might disagree, heh) and Xeon is perfectly fine for small jobs, but when you want High Performance Computing, step into the Opteron 64 realm.

As for Wusy w/ DHCP, of the over 100 Windows Server 2003 Enterprise Edition machines I've built to run DHCP, I have never seen a faster HDD be the deciding factor in the performance of that server. I've ran them on systems from P3 800MHz to a P4 3.4GHz and a Dual AMD Opteron 2.0GHz w/ 4 Western Digital Raptor 74GB in RAID 0, being that DHCP simply leases out IP's (and hosts of more crap, darn DHCP!) HDD speed is not vital, if not null. But I do agree, VIA is from far depths of Hell. (I have ran it on a P2 333MHz w/ 5 18GB 10,000RPM SCSI in RAID 0, but we won't go there....)

For a recommended setup: 4 Servers
Server1: DHCP, DNS & Print Server
Server2: Financial Apps & Records Management (You want a Dual CPU for this one)
Server3: File Server & Mail Server
Server4: Terminal Services

Server1 can be a low-Athlon XP or Pentium 3-Low P4, 512MB of RAM is plenty for your situation.

Server2 should be your Dual Xeon 3GHz you have right now, that 2GB RAM is perfect.

Server3 should be a Single AMD Opteron 252 or 254 (2.6GHz/2.8GHz) with 512MB to 1GB of RAM or more and RAID5 (For Speed/Redundency) of Fast SCSI or Raptor 74GB HDD's.

Server4 should be the Dual AMD Opteron 248 or 250 (2.2GHz/2.4GHz) with 2GB or more of RAM (Being Terminal Services can rape a Server)

1 Thing to note, at least in my experience, is that Running x64 Versions of any OS (XP Pro x64, Server 03 x64, Linux x86_64) will greatly increase performance of that system, I have seen up to 20% or more performance increases running x64 vs. x86.

EDIT: I wouldn't agree with Stimpy on Clustering, my experience in Clustering (From 25 P2's with NT 4 to 8 3.4GHz Server 03 machines) it has been a disaster, I would only recommend clustering w/ Linux (worked perfectly and much better performer than NT 4 or Server 03) at least in my experience.

~~Mad Mod Mike, pimpin' the world 1 rig at a time
February 8, 2006 11:30:26 AM

Both Opterons and Xeons will do fine for the job but Xeons might be able to handle the workload better, thanks to their HT advantage (which, obviously, is worth nothing if the application doesnt support it, then the Opterons might have the edge).
Well... what it all comes down to is costs and stability and this is where Xeon has the biggest advantage. Face it, nothing can beat CPU + Chipset + MoBo + graphix all by the same company and even in case that system fails, its way easier to get everything done if you only have to contact 1 company.
February 8, 2006 12:51:00 PM

Being an Intel fan i hope that the above post is corect.
Anyway, even if they aren't so good as servers,opterons still rule.
Is that an EE? Is that a FX? NO,it's an opteron OCed at 2.8 STABLE :lol:  .
February 8, 2006 1:21:57 PM

I agree with your statements and also wanted to add to the Dell information you pointed out. Dell is moving towards AMD because they want to be able to control manufacturing costs better. Being the largest computer manufacturer, they can do this by having the two processor brands compete which usually drives prices lower.

Also, Dell wants diversity. By offering AMD processors, that will certainly appeal to the gaming crowd and that's where Dell wants to go.

One more thing, I read the part about all those machines you listed in the one post. I would sure hate to have to pay your power bill .. lol 8O 8O 8O :D  :wink:
February 8, 2006 5:37:02 PM

Quote:
Both Opterons and Xeons will do fine for the job but Xeons might be able to handle the workload better, thanks to their HT advantage (which, obviously, is worth nothing if the application doesnt support it, then the Opterons might have the edge).
Well... what it all comes down to is costs and stability and this is where Xeon has the biggest advantage. Face it, nothing can beat CPU + Chipset + MoBo + graphix all by the same company and even in case that system fails, its way easier to get everything done if you only have to contact 1 company.


Whatever you're smokin', I want some of it.

~~Mad Mod Mike, pimpin' the world 1 rig at a time
February 8, 2006 8:06:30 PM

If you are referring to Intel HT (hyperthreading) it is actually a liability not an advantage and virtually all Intel CPU are benchmarked with hyperthreading turned OFF.

This is what Intel has decreed.

In server usage hyperthreading can actually degrade performance instead of helping... funny thing huh??????
February 8, 2006 8:14:06 PM

Quote:

Well... what it all comes down to is costs and stability and this is where Xeon has the biggest advantage. Face it, nothing can beat CPU + Chipset + MoBo + graphix all by the same company and even in case that system fails, its way easier to get everything done if you only have to contact 1 company.



I respectfully disagree.

Especially on the Intel side Intel CPU + Intel Chipset + Intel MoBo + Intel GFX = Junk

Intel Extreme GFX is ANYTHING BUT extreme, my 20 year old ATI Mach 64 beats the crap out of it oh and it's ISA :D 

Intel Chipsets are usually known for slower perfermance. 3rd party chipsets for Intel CPUs usually perform a lot better.

Has everyone forgotten the Pentium bugs, RDRAM and i840 chipset fiascos among others?
February 8, 2006 8:16:13 PM

As one that usually is a little more Intel biased, I have to disagree with you there man. The HT for server applications just doesn't cut it and as linux said, it can actually cut down on performance. Intel has Conroe coming out for desktop replacements as well as an entirely new server chip. Intel knows their chips don't cut it in real world server apps so that's why they're going back to the drawing board.

Most people will tell you that in the desktop world, Intel doesn't cut it either however, if they're given the right environment, they will work just fine.
February 8, 2006 8:17:25 PM

Thanks :D 

I agree with you too ;-)

Those machines are not all on at the same time so my pwr bill is actually very reasonable.
February 9, 2006 5:12:10 AM

Quote:

I work for a small local government council. We have about 70 users total.
30 - 40 Terminal server.
40 - 30 Local Network.


Gotcha.

Quote:

Currently we have only ONE server. Which handles everything, DNS, DHCP, Finance Apps (Interbase), Records Management (SQL), Print Server, File Server, Mail Server, and the real hardware killer is the Terminal Server (25 users average. 40 users Peak)


Ouch.

Split that load up among multiple servers, NOW. Everyone here is telling you to split it up, and I'm starting out by telling you the same thing. Maybe if you show your bosses all the posts telling you to split it up, they'll realize just how thick-headed they're being about this.

Put DNS and DHCP on one server. An old piece of crap P3 or something is fine, as long as it's stable and has a fair amount of RAM. DNS (BIND in particular) keeps all its records in system memory and commits very little to disk. Putting DNS and DHCP on the same machine has the added benefit that you can keep DNS records in sync with DHCP assignments without exposing more services on your LAN. Windows is overkill; use Linux or BSD.

Print and file services can go on yet another machine. Same requirements as your DNS server, except you probably also want fast disk access. You can probably put mail services on there too, if you beef up the processing power and memory. In fact, I'd probably leave this to your current Xeon setup. Windows is overkill; use Linux or BSD.

Interbase and SQL services should go on a dedicated machine. 2-way or higher Opteron is highly recommended, especially for database services, especially if your database software has an x86-64 version available. Databases typically get HUGE benefits from going 64-bit, and Intel's x86-64 implementation (EM64T) frankly sucks wind. You'll need fast disks and lots of RAM. Use whatever OS supports your software.

Terminal server, again, dedicated machine. You should expect that your terminal server will get compromised occasionally, so don't run any other important services on it. Use Windows if your users demand a Windows interface on the terminal server; use Linux or BSD otherwise. You might as well run a 2-way or higher Opteron server. Expect most of its file I/O to come off a shared filesystem on your file/print server, so blazing-fast disks aren't a big priority.
February 9, 2006 6:12:40 AM

Quote:
Both Opterons and Xeons will do fine for the job but Xeons might be able to handle the workload better, thanks to their HT advantage (which, obviously, is worth nothing if the application doesnt support it, then the Opterons might have the edge).
Well... what it all comes down to is costs and stability and this is where Xeon has the biggest advantage.


Completely disagree.

Without Hyper-threading, Opteron still rules the server market that Dell needs to cut down the prices of their Xeon Ovens.

Quote:
Face it, nothing can beat CPU + Chipset + MoBo + graphix all by the same company and even in case that system fails, its way easier to get everything done if you only have to contact 1 company.

Only if you buy the motherboards, ram, hdd, etc. from Intel.
February 9, 2006 6:46:23 AM

Quote:

I'm in Australia so we don't have NewEgg here, sadly.


EDIT: HyperThreading is not application specific, or supported, it is either ON or OFF. An application that does not execute in 2/4 threads, and heavily use SSE2 (eg: SSL, Possibly SHA512, MD5, etc) will not gain anything with HT enabled, and in many cases throughput might rise a tiny bit, but response times will increase too unacceptable levels. Dual-Core (be it a cheap Pentium-D server on a decent server chipset, or a 4-Way/Socket Opteron with 8 processor cores and 4 way NUMA) processors are a far better choice for these reasons. (Summary only, é-mail contact on my webpage - See signature for URL).

Intel have already abandoned Hyper-Threading in the server space. Get with the program ! , Perhaps if they try for 'HyperThreading II' it won't be so bad in 2-3 years. But Intel are currently aiming for 4-issue server processors with 2/4 cores, a far better approach than 2-issue processors with HyperThreading ever was, to be released in Q3/Q4 2006.

Above had to be said, too much 'disinformation' during the 'information age' ironically. (That or kids forgot how to read / understand techdocs & whitepages years ago - ???).
================================================

I too am in Australia.
http://users.on.net/~darkpeace - is my dodgy ass website with my contact details.

You might want to consider something from Sun Microsystems, http://www.sun.com - they have a decent footing in Australia. They offer x64 AMD Opteron servers.

If you can support it yourself then grab a Tyan system - http://www.tyan.com - SuperMicro, etc are not 'that easy to find' in Australia, Tyan gear is just as nice to work with.

Remember my contact details are on my (dodgy ass) website.

HyperThreading is good for servers performing lots of SSE2 based stuff, like encryption / decryption, etc. It only raises MIPS by about 20% or so, but bear in mind that means 2 threads run at 60% performance instead of 50%, not that 1 thread runs at 120% performance (duh, but needs to be said as so many 'admins' have no clue about hardware, they won't admit it though but wouldn't last 15 min in a room with me :p  .... as you likely are already aware.)

The Opterons will be cheaper in the long run to run, over 2-3 years. (Look at Sun Microsystems website). Most people here will agree with their findings. Sun also make multi-threaded UltraSPARC processors, mostly for Solaris / Unix OS, but the documention may be of 'some use' when 'studying' Intel HyperThreading (I find many people don't quite understand how it works, and expect higher performance than it offers.... it actually decreases memory performance slightly when enabled due to 'sharing overheads' as there are more logical processors trying to access memory at once, and can decrease performance in various tasks when enabled... aswell a increase, but only 10-20% in server apps... Usually you won't hit 100% CPU load on both 'threads' either, as most software isn't designed for it, 50% on one both threads means it is max'd out, any more indicates the software being run is likely taking advantage of it.) - It is hard to go into depth on it in just one post in a forum. (You need to certify applications for the environment one by one, see if it helps or not, various CPU internals are only half effective when HT is enabled and the software isn't benefiting from it, which is why disabling HT can give better performance than just specifying affinity of each process, etc, etc, etc - Contact Me).

If you have RealVNC I'll be more than happy to let you remote desktop my rig here, (I also work in Gov, and we need to help each other, the stuff I see every day 'proves' it - hehehe) and get a feel for it if your are interested - Both in Australia so decent latency, etc, I also am not working this Friday (10/02/2006) so should be available on MSN. (Rig specs below, but can be respec'd more as a server with 3 x PCI-X slots (for RAID cards, etc), and 2 x PCIe slots (both of which can be used as has cheap ATI Rage XL onboard, since servers don't need 3D performance).

I've got SANDRA, but using RAID-0 on 7,200rpm HDDs (ample performance for what you may be doing though, the SCSI is optional and not always justified now that SATA HDDs have NCQ and large caches, P2P interfaces, etc... SATA/SAS is technically superior to older 'ribben' SCSI, people will argue otherwise, but they are just in denial most likely.)

The stuff you can build yourself & support with fast turn arounds, vs buying gear from Dell, etc may actually justify in 'in house' custom server(s) for this project, as we are only talking about 100 staff or so. Especially if there are support staff near the site, and the saved cash is put into hot spare equipment. (Beats waiting an hour or two for '3rd party' support, while business suffers, or is badly impacted, etc).

The whole 'rig' would cost under AU$10,000 , likely well under it too, or get 2 'more mid-range' rigs for a similar price... and remember to keep spare equipment in case of failures.... Also suggest a LSI or Adaptec, etc PCI-X RAID-5/6 controller for this purpose. (and no 3D card obviously enough :p , you will be able to use both the PCI-X (dedicated briges too) and PCIe x16 for RAID controllers in the future aswell. (assuming you are considering a Tyan K8WE S2895 in house build.

All my original notes are at:
http://users.on.net/~darkpeace/hardware/Opteron270.html
Get 128x4 RAM aswell, not 64x8 RAM, as ChipKill(tm) benefits only work on the 128x4 stuff... sure it costs more but stability over a 2 year period is important... why settle for single bit correction and dual bit detection, when you can have 'almost 99.99%' uptime without memory issues ? - In your case, and mine, the extra few dollars justifies the reliability. (Does Wiki have a entry for ChipKill(tm) ?)

Contact Info on my website - See signature below.
February 9, 2006 7:21:30 AM

Quote:

I also wish to go Opteron, however the partners I'm dealing with view AMD as a poor unreliable company (VERY hard to pursued).


Just get RealVNC and connect to my server for a demonstration.

www.top500.org - How many Xeons vs Opterons do you see the Top500 using ?, I mean these old ones need to wake the hell up and smell the coffee... ignorance is not an excuse, how many Australian IT projects have failed in the last 6 months alone.... Do they really want their names added to the list of growing failures ? (hell, they prob do). Just look at the problems Customs had right before Christmas (really picked a good time didn't they ?). The cause was underpowered systems, they heavily overestimated what could be done, knew the risks, and did it anyway. (The one person at fault was advised it 'simply would not work' 100's of times, but did it anyway).

HDD performance isn't my rigs strongest point (see sig), but I am beating the iRAM drives in 'some' areas.. not seek time obviously using a Tyan K8WE S2895 with Opteron 270's (Don't bother with the 250's and single core stuff, get maximum benefit today, x64 options for the future, sitting there ready to go, with 2 processor cores per socket). Having the option to move up to 8 GB Reg ECC PC3200 (128x4), using NUMA (built into Windows XP Pro, x64 Edition and server editions) will help in terminal server heaps.

Not that Intels Xeons are bad, but we've had enough issues with them anyway (Mostly people thinking HyperThreading = double performance :p , sad but very true, most the stuff we do it is better off disabled - I have a long document explaining why too, in great detail exactly what happens, just mail me). The Quad-Opteron is my 'home PC' (hehehe), just wanted to hammer them and check for any differences. :p  , Intel need to start pushing those Xeon LV's really soon, they've lost significant market share to AMDs' Opterons in since around Q4 2003).

It is all about non-biased technical documentation, and using whatever performs the best, with the best price/performance ratio, power consumption, TCO, reliability, etc for the tasks at hand... I can help you here.

Note: You are not going to be using Ghost, or 'typical' imaging on the server(s) anyway because of the RAID-5/6 controllers, (it won't work, use their imaging / rebuild software), and that is the only real reason I continue to keep a few Intel systems around. (Intel IDE controllers work better with Ghost in a DOS env, that is about it).

For workstations in a work-place, go Intel, there are advantages, but for the heavy-metal where performance is required and IDE controllers are not even being used the same excuses don't really work do they ? :p 

The reason some admins fear anything but Intel is because they don't want to look like a dick, as they lack experience in everything but Intel CPUs, platforms, chipsets, etc.... they also fear being replaced by younger / smarter people.... it is a very lame excuse, they are only like half a tech when you think about it. (Become familar with ALL hardware computing platforms you lay your eyes on, and some you don't.... that is the best advice I can give you to help your workplace move to a new level, possibly with you at the helm one day).

Also: Real Australians use http://www.StaticIce.Com.Au ;) 
February 9, 2006 7:56:21 AM

China's 2nd biggest Cluster is Opteron based.

Some of the largest military, scientific and academic users in the US and around the world use Opterons.

As I stated above at LEAST 45 of the top 500 supercomputers on earth are Opteron based machines and the vast majority of the top 500 run Linux, BSD or Unix. Almost none run windoze.

This is despite the fact Intel has a virtual monopoly and has been involved in illegal anti-competitive behavior for decades.

Penguin Computing, HP/Compaq, IBM, Sun and others all make Opteron systems.

Please note that HP makes AMD systems despite the fact they compete against Intel which happens to be their partner in developing the Itanium processor which is commendable. Also HP has been pushing Linux and I applaud them for that.

I would recommend custom built machines if that is possible, otherwise get them from a major vendor like Penguin Computing, HP/Compaq, IBM, Sun, etc.
February 9, 2006 7:56:22 AM

I have to disagree with "HT is a minus".
In most cases HT brings you a lot more performance. Especially in multitasking. It fails at DiVx but at most other tasks it is at least the equivalent of a non HT CPU running at the same clock.
The pros heavily outweigh the cons.
HT isn't the inovation of the 21 century but it certainly makes a difference. In most cases an improvement.
February 9, 2006 8:15:28 AM

Quote:
I have to disagree with "HT is a minus".
In most cases HT brings you a lot more performance. Especially in multitasking. It fails at DiVx but at most other tasks it is at least the equivalent of a non HT CPU running at the same clock.
The pros heavily outweigh the cons.
HT isn't the inovation of the 21 century but it certainly makes a difference. In most cases an improvement.


Mate you are ****ing clueless aren't you ?

Where do I start ?:
- "HT brings you alot more performance in most cases", no you gain about +20% thoughput at the cost of (possibly getting up to) double the response times. In SSE2 you may gain up to +99% (double) though, but such cases are 'extremely rare' not 'most cases', not even close.
- HT 'fails' at DivX - No it doesn't, it performs about +85% better when used.
- The pros heavily outweigh the cons (half truth, can argure either way)
- Server tasks are usually memory performance dependant, the CPUs would ideally sit at 30% load (average over 6 hours), and memory is the main 'bottleneck', HyperThreading hurts memory performance (read a TechDoc or WhitePaper one day and get a clue), add to this AMD has a NUMA capable platform now which can aggregate their memory performance (each processor acts like a northbridge to the other) to 12.8 GB/sec or 25.6 GB/sec (peaks yes, but still well over the Xeons which typically have 6.4 GB/sec, peak, being shared by 4 logical processors which adds overheads.)
- We are not talking about video encoding here, (which is surprising to some, does not actually scale up/down with memory performance, it is mostly CPU and registers, HT actually helps here, you claim it doesn't ?) - (Sheesh, Idiot, Everyone here can show you HT helping DivX, care to RealVNC someones box ?), nor are we talking about physics engines in games, or machines that are performing mostly authentication (sure they do some, but not 'mostly'), which, as en/decryption benefit from HT/High SSE2 performance, would perform better with HT on.

Sun Microsystems implementation of MultiThreading was far more thought out. Go read about it, should keep you busy and educated for about 2 weeks or so.

Also when moving to an x64 server operating system, the Opterons gain more in Drystone / MIPS than the Xeons do.

H/T doesn't run one thread +20% faster, it runs two threads but intead of being executed at half-speed, they execute +20% faster than half speed. Thus the appearances. Anyone dealing with real-world applications knows this already.

http://www.xbitlabs.com/articles/cpu/display/replay.htm...
http://www.xbitlabs.com/articles/cpu/display/netburst-2...
Get reading, and may I suggest get learning aswell. :p 

Also bear in mind that H/T has security flaws:
http://www.daemonology.net/hyperthreading-considered-ha...
http://www.daemonology.net/papers/htt.pdf
Now I know you may like backdoors into Terminal Servers, but the admins don't want you there remember ?
February 9, 2006 8:25:52 AM

Very eloquently put :D 

I agree!

It appears we were both typing at the same time.
February 9, 2006 8:32:40 AM

Quote:

Intel CPUs score a bit better with Hyperthreading turned OFF.


Should be edited to:
"Intel CPUs score a bit better with Hyperthreading turned OFF when used in most server based applications / roles... a minority of cases do exist where it helps. eg: a server performing almost only authentication tasks, where throughput of logins is more important than response".

Just for that 'all inclusive' finishing touch ;) 

============================================
Now back to notes:

One must bear in mind almost no servers in the original post are performing 'almost exclusively' authentication based tasks. The Terminal server is a so-so grey area..... However, an Opteron 270 would be a far better investment and last quite some time.... it appears this 'department' doesn't upgrade nearly often enough so it is justified TCO wise, as it may very well 'keep performing well enough for even more users' for 1-2 additional years, down the track, than it is required to, even if they 'forget' to upgrade it every 2 years or so to keep it 'scaling with demand'.

If you have 100 users, try and plan for at least 150, if not 300. That way you save your own ass, and Government jobs are all about 'saving your own ass' these days. (At least in Australia).

Least if your managers screw it up, it is them that are liable, not you. If it works (well regardless really, keep your own notes) document the project and add a summary to your resumé, so if you need to leave in a hurry (eg: dumb project coming up and they want you as a scapegoat) you are well prepared for it.
February 9, 2006 8:59:42 AM

I went ahead and changed it to:

Intel CPUs score a bit better with Hyperthreading turned OFF in most server tasks.

Which should cover it.

Quote:

However, an Opteron 270 would be a far better investment and last quite some time....


I totally agree with your comment about the Opteron being a much better investment.

Furthermore I would argue that AMD CPUs in conjunction with Linux, BSD and other open source software would offer a MUCH better TCO and would provide a lot more value to Australian tax payers!


I believe the OPs superiors would be in a lot of trouble with the taxpaying voters in Australia if the public were to find out the organization could have saved tens of thousands of AU dollars by going with AMD and open source.
February 9, 2006 9:04:31 AM

Why don't you get some reading
http://www.2cpu.com/articles/43_3.html
I have HT P4 and i probably know what i'm saying. I don't need benchmarks. Benchmarks are a crap way to test something and in some you'll find X>Y and in others X<Y.
I don't do server applications but i don't see why having an extra fake core wouldn't help.
For me HT on means
better multi tasking
better performance in certain games
And Ht off means
better performance in other games
usual single threaded apps it doesn't make much of a difference(and the difference is 20% at MOST).
Stop basing on facts that come out off your small mind. We all are good at theory but when it comes to something practical things change.
And HT can be turned off whenever needed (THAT makes it an Improvement).
As for servers keep whining LOL, the author's boss doesn't trust AMD thus will go for Intel.
EDIT
After further reading it seems that HT brings no improvement at servers.
I admit it's nice that you pointed that out.

P.S. Thx for calling me Idiot, fact based freak ^**%.
February 9, 2006 9:08:59 AM

and, as likely above:

- Replacing / Fixing hardware using staff in the same building if they do it themselves. vs. Having to wait... sometimes quite some time.

- Why wait 2-4 hours for '3rd party OEM company tech' (eg) to replace a HDD in a failed array, when you can do it yourselves faster, and at a lower cost. It takes 5 minutes and is the piss easiest 'preventative fix' you'll ever perform on a server. (The drives will fail, eventually, just have 'tested each quater' ones on hand to use as hotspares).

- Can you really be without a server for 2 hours ?, While simply waiting for '3rd party OEM company tech' to arrive, just to perform 5 minutes of work ?

- Same applies for data backups, mirroring, (clustering if used, etc), How long can a given 'server / backend' core business machine (S/W or H/W) be out of action before money is really getting burned paying 100 (?) people to 'wait for the servers to come back online with all their data' ?

(I think wusy, or someone else, may have covered this already in more depth, call this a 'page 2 reminder')
February 9, 2006 9:17:43 AM

I agree and would advise every organization to keep spare hardware on site whenever possible!

I speak from experience, having to wait for a vendor to drive to your location or having to wait days for a part to be shipped can cost a lot of time, money, sleep, peace of mind and jeopardize your job security.

This is yet another reason why you should NOT run all those critical services on a single box as many of us have pointed out above.
February 9, 2006 9:24:22 AM

Quote:
Why don't you get some reading
http://www.2cpu.com/articles/43_3.html
I have HT P4 and i probably know what i'm saying. I don't need benchmarks. Benchmarks are a crap way to test something and in some you'll find X>Y and in others X<Y.
I don't do server applications but i don't see why having an extra fake core wouldn't help.
For me HT on means
better multi tasking
better performance in certain games
And Ht off means
better performance in other games
usual single threaded apps it doesn't make much of a difference(and the difference is 20% at MOST).
Stop basing on facts that come out off your small mind. We all are good at theory but when it comes to something practical things change.
And HT can be turned off whenever needed (THAT makes it an Improvement).
As for servers keep whining LOL, the author's boss doesn't trust AMD thus will go for Intel.
EDIT
After further reading it seems that HT brings no improvement at servers.
I admit it's nice that you pointed that out.

P.S. Thx for calling me Idiot, fact based freak ^**%.


Note, the below comments are not directed at you, just notes to assist anyone skimming this forum thread for insight into the situation at hand. - It was just easier to use your above post, to save time / as a reference, even though it says I am replying to you... it is more of a general reply to some of the issues present.

Each bolded section is given a number, from 1, counting upward.

1 - What only probably ?, we are talking Xeons here anyway, likely the older 533 FSB model ones aswell, all sharing that 4.3 GB/sec :p 

2 - We are doing server apps, and sometimes it will help, and sometimes it won't. Use the right benchmark for the task, Thus you don't get any X>Y then Y>X scenarios, you've planned for them all.

3 - Exactly, but servers may be running the same single theaded process, and have it loaded and executing on 4 processors (logical or otherwise), but with HT ON, they max the chips and each 'instance' of the code slows to 60% performance, you get thoughput at the cost of response in most server applications when HT is ON. (As many have said above). Like I said +20%, and +20% of 'half'(or 50%) (what one would normally expect when executing 2 processes on one CPU with HT OFF) is equal to 60% :p  (best case as you've said, so it is even less effective 'typically').

4 - Why the hell would anyone base this on FICTIONAL information ? (BannanaBoat. laaalaaalaaa.....) (Best comment ever)

5 - I wish my mind was small, better than dealing with ignorance on a daily basis, aswell as on the TG forums (the last place I'd expect to see it).... Ignorance to you could well be bliss, thx for the warning ;)  (I am kidding of course, our posts compliment each others)

6 - If ones practical and theoretical scenarios are differing, then the theory is wrong or 'an imcomplete scaled down version for simplicity', this is normal in the business world..... the real-world.

7 - You can't just go rebooting servers every 2-24 hours to toggle HT on and off for the task at hand. Normally servers are managed remotely and you can't even get into the BIOS to turn HT on/off. Using process affinity locks doesn't help either (prefetching, OoO, register renaming can't look as many instructions ahead, only 40 instead of 80, or only 80 instead of 160, either way you may aswell have just left HT off). --- Sort of a flipside comment, did you change your mind while writing your post ?

8 - I do sincerely apologise for calling you an idiot though (above). You ain't so bad, just more that I am on a personal crusade to turn Australian Government IT around, for the better - :)  . I did underestimate you based on the post alone, and that was a mistake on my part ;) 

Please don't take offense at this post, you've helped us both point out a few things often omitted.

Also note that your linked article "Hyper-Threading Performance Analysis - DivX and Conclusion - Published on 2002-09-30 13:57:42 By: Jim_) was... for starters created by someone simply called 'Jim_' on the internet :p , it is also using DivX software from prior to Sept 2002... The current versions (we are Feb 2006 now, almost 30 months from the article publish date and 'old version' software used) of most video encoders are actually gaining in HyperThreading, WME9 x64 (4 isolated threads) has been out for a month... WME9 (x86/32) (2 isolated threads) even longer still.... but yes a upper mid range Opteron, even a single core one, can encode video, even in when only x86 32 bit Protected Mode, just as fast as a high clocked Pentium 4 / Xeon with HT on or off.


Note: It appears we are actually in agreement on several points.
February 9, 2006 9:50:15 AM

Probably=Certainly in that context
Settled.
Linux,stop trying so hard,Intel is the only way. ><
"I also wish to go Opteron, however the partners I'm dealing with view AMD as a poor unreliable company "
^^
I'm so Intel BIAS i can't stand it . :lol: 
February 9, 2006 11:02:25 AM

Not to take away or prove any one wrong...but take a look at this url

http://www.microsoft.com/technet/prodtechnol/sql/2000/p...

About the second paragraph, you'll read this:

"To date, the largest configurations supported by SQL Server 2000 running on Windows Server 2003 are 32 processors with 512 GB of memory. These systems demonstrate excellent SMP scalability, both on audited benchmarks and in real applications. Today, a single CPU can support 14,000 users accessing a 1-terabyte database, an 8-processor node can support more than 92,000 concurrent users accessing a SQL Server managing billions of records on a 8-terabyte disk array, and a 32-CPU node can support 290,000 users accessing a SQL Server database hosted on 24-terabyte disk array. The largest of these servers are capable of processing more than 1 billion business transactions per day."

"Today, a single CPU can support 14,000 users accessing a 1-terabyte database"

Now of course, your nowhere near those numbers on networking and I/O's... but it gives you the idea of what a cpu can handle...which is [IMO] not your problem...

You didn't say [or I didn't see] what OS your running... but when your installing Windows 2003 Server, they give you the option to install all the stuff that your running on one system, so sitting there and saying that you need 3 or 4 servers [IMO] No... why would Microsoft give you the option to put them on one system if it's not designed for it.... [Giving your size]...meaning if you had a 1000 users... then of course... utilize another server for certain things... but in your case...No.

I also see you have 2gb of RAM installed... I recently read about a Microsoft/H.P. study on the amounts of RAM utilized when using "Terminal Servers"... and in the report they say that each connection uses about 8 to 12 mb's of RAM and can go up to and over 50 mb's... depending on whats being used.

The url below is another study/report that will give you the same numbers

http://www.sessioncomputing.com/scaling.htm

about 85% of the way down, you'll see this answer:

"256MB for Windows Server + 10-15MB for each logged on user + RAM required by applications, i.e. a user can easily consume 40-50MB or RAM by running Outlook, Word, Excel and Internet Explorer simultaneously. Since many applications use dynamic memory allocation and share memory, you can't just multiply the number of users times the amount of RAM one user consumes. I generally recommend 30-40MB per user as a baseline for users of Microsoft Office."


You might want to think about updating your Server OS... if not running 2003 or maybe your system needs a service.

I would also look into updating your drives to a "raid" system, for quicker I/O's...

but as far as what type of CPU...I wouldn't worry... your way above and beyond what you'll need....

IMHO
February 9, 2006 11:27:32 AM

Nice articles btw:

Depends on what the SQL is used for (as we both know).

Those 'best case' examples indicate a 'average' user can only perform 20 transactions per hour... (and you need to plan for your heavier, more productive users else they complain about performance).

Now in hard working places people tend to do more than one transaction every 3 minutes, they could be doing 1 every 5 seconds.... or every 15-30 seconds in reality.

100 users - 2,000 queries per hour could very well actually be:
100 users - Up to 36,000 queries per hour. or:
50 users - Up to 18,000 queries per hour. etc...

Check: Are your users still complaining about slow response ?, if so then do something as adding staff will only make it slower, adding hardware may be the cheapest way to increase productivity.

Averages are dangerous things, as they exclude spikes, and the spike load may be a 40x heavier load (on some components) than the 'average'. Perhaps these averages included idle time, lunch breaks, etc. :p 

It totally depends what the SQL is being used for, how 'fast' the users work, how well the software they are using which accesses the SQL is designed & configured, coded, etc.

Wanting 6-8 GB on a terminal server (or split over two terminal servers rather) for performance terminal sessions wouldn't be out of the realm of possibilities here either, it depends which applications staff are running, how memory heavy they are, etc, even with only 50 sessions. You need to plan 'against this' so the system isn't 'slow' by the day it is implemented.... and hopefully runs fine for at least 2 years afterwards... Perhaps staff are also using applications that burn resources (there are many around, thanks to IT booms, poor devs for the time, and low quality software rush jobs)... If they are just using e-mail and word, and nothing else then yeah it is enough.... also depends how fast the sessions are connected, the slower they are the less they can hit the server... (which is a double edged sword productivity wise anyway).

Sure anything would be an improvement on their current server though.

Once all the infrastrucuture is 'cool'... The biggest improvements can be only really be made in the software the users are running most likely, eg: better designed, doesn't 'machine gun' the SQL Server, etc... getting developers to fix code on the other hand... is that even possible ?

The people who made it could've:
- coded it 4 years ago,
- pissed off never to be seen again,
- likley where paid a large sum of money for something that is may be 'designed' to 'keep them employed on an ongoing basis' (hint, hint, look here).

I've been around here long enough to see all the above (sadly :(  )
(Note:: Poster & myself both work in Australian IT, in Gov sections, likely they are dealing with similar, if not the same, problems we've dealt with before.)
February 9, 2006 1:00:15 PM

Quote:

Also note that your linked article "Hyper-Threading Performance Analysis - DivX and Conclusion - Published on 2002-09-30 13:57:42 By: Jim_) was... for starters created by someone simply called 'Jim_' on the internet :p , it is also using DivX software from prior to Sept 2002... The current versions (we are Feb 2006 now, almost 30 months from the article publish date and 'old version' software used) of most video encoders are actually gaining in HyperThreading, WME9 x64 (4 isolated threads) has been out for a month... WME9 (x86/32) (2 isolated threads) even longer still.... but yes a upper mid range Opteron, even a single core one, can encode video, even in when only x86 32 bit Protected Mode, just as fast as a high clocked Pentium 4 / Xeon with HT on or off.


Note: It appears we are actually in agreement on several points.

LOL,you aren't finding anything based on the "new" version on the Internet. searched myself,everything dates back before 2003.
And having a HT P4(crappy 2.4Ghz) i wished it did offer better but it doesn't. Probably my old progs are outdated (and they are,i only update NAV and gfx drivers+games) but heck, this CPU is outdated as well.
So lets see what's on the list Intel 650(?) 3.6 ghz CPU with HT + arctic freezer of course(or something better,i don't want it to keep me warm during the summer :? ). Then i'll redo the DivX tests with updated software. Until then( and it'll take a while since i'm starting this rig from 0 and i need more $$$) i remain to be convinced 8) .
February 9, 2006 5:55:41 PM

while HT might not be best in a server situation (it can be disabled), I have found a definite speed increase both in normal every day life, and in benchmarks by having HTT turned on. Ripping CD's with Windows Media Player uses two threads... something very rare. Usually when I run benchmarks, I only use about 55% CPU, because one thread is taking 50%, and the other side is free for everything else, but when I use WMP to rip CD's to MP3, it uses 100%. If you have apps that take advantage of HTT, you have a definite advantage. :) 
February 9, 2006 8:07:28 PM

I agree with you there Windshear and your right. HT does make a noticable difference in the desktop world. Of course, the AMD X2 is pretty impressive I must say. I'm really into this new machine of mine. Could this be the end of Intel for Luminaris? hhhmmmm :?

Now in the server market, that's a different ballgame my friend. AMD just shines and I know that from experience.
February 9, 2006 10:42:56 PM

Have you ever considered SUN servers??

They're excellent performers and are very affordable (you can buy one starting at $745 u.s). They even come with Solaris 10, Red Hat or even Windows Server 2003.

Take a look here and here

You can show the benchmarks to your boss and maybe he'll change his mind.

(SUN doesn't use Intel processors. You must guess why?) :wink:
!