Quad Core Intel Xeon E5410 Harpertown 2.33GHz = equivalent to what AMD CPU?

[Trizik]

i'm trying to squeeze everything into a single 1U box to keep my colo expenses down

i'll be running 8-10 game servers (mostly srcds): the affinity of each large server (32 slots) will be set to a different core, the affinity of a couple smaller servers (20 slots + 10 slots) will be set to the same core

in my past experience hyper-threading had such a negative affect on the performance of these servers that when they were near full the in-game choppiness was unplayable. disabling HT resulted in smooth, playable full servers. from what i've gathered across various forums, that still seems to be the case with this specific game server software. as such, i am not interested in HT. if you want to convince me otherwise, please try because i am all ears

i pieced together two systems; same case, same hard drive, different mobo, different CPU, different RAM:

2 x quad core Harpertown 2.33GHz 8GB DDR2 = $1,451.91

2 x quad core Gainestown (Nehalem) 2.0GHz 9GB DDR3 triple channel = $1,441.91

the Nehalem-based system is pricing out $10 cheaper than the Harpertown system. i just need to be sure that a single core of the 2.0GHz Nehalem system is indeed more powerful than a single core of the 2.33GHz Harpertown system because i am sure a single core of that Harpertown system can run a 32 slot server just fine and dandy all by itself

my budget is $1,500

 

[Dadofamunky]

wont fit in standard boards.

Follow the thread..
http://www.xtremesystems.org/forums/showthread.php?t=237196

But a 4x4 AMD setup would be pretty badass, and you can get them for cheap if you look around.

I dunno, when it gets down to soldering power connections on the mobo that's getting a little overboard... mucking around with old motherboards and such for a pretty modern app like that seems to have diminishing returns to me. I think A and others are right - think seriously about an i920 build. If a 4x4 AMD setup could be had without soldering traces on the mobo, and you can save money, then sure, why not

 

[jones377]

i'm trying to squeeze everything into a single 1U box to keep my colo expenses down

i'll be running 8-10 game servers (mostly srcds): the affinity of each large server (32 slots) will be set to a different core, the affinity of a couple smaller servers (20 slots + 10 slots) will be set to the same core

in my past experience hyper-threading had such a negative affect on the performance of these servers that when they were near full the in-game choppiness was unplayable. disabling HT resulted in smooth, playable full servers. from what i've gathered across various forums, that still seems to be the case with this specific game server software. as such, i am not interested in HT. if you want to convince me otherwise, please try because i am all ears

i pieced together two systems; same case, same hard drive, different mobo, different CPU, different RAM:

2 x quad core Harpertown 2.33GHz 8GB DDR2 = $1,451.91

2 x quad core Gainestown (Nehalem) 2.0GHz 9GB DDR3 triple channel = $1,441.91

the Nehalem-based system is pricing out $10 cheaper than the Harpertown system. i just need to be sure that a single core of the 2.0GHz Nehalem system is indeed more powerful than a single core of the 2.33GHz Harpertown system because i am sure a single core of that Harpertown system can run a 32 slot server just fine and dandy all by itself

my budget is $1,500

Running 8 single-threaded applications on an 8-core server is not a whole lot different than running 1 application that supports 8 threads. If you would run just 1 gameserver on these systems, then the Harperton may be about as fast as the Gainestown and alot faster than most Opterons. But by running 8 of these at the same time you put ALOT more pressure on the memory system, resulting in a slowdown for each gameserver. This slowdown is less on Opterons and Nehalems than on the old FSB Xeons due to their superior memory subsystem. They can also take fully advantage of NUMA optimisations in the OS because each gameserver is using it's own pool of memory.

I can use an example with Folding@Home. Granted I have not kept up with it the past year but lets assume that F@H is still singlethreaded for the purpose of this discussion. When multicore processors arrived, people who wanted to fold at max speed simply ran one F@H process on each core in their system.

Now lets say we have an hypothetical single-core CPU at 2GHz running F@H. And lets say each Work Unit (WU) takes 3 hours to complete. Running it 24/7 and you would average 8 WUs each day. Now we have a dual socket CPU, with 1 cores per socket for 2 in total. Because the second socket adds coherency traffic and increases memory latency a bit we may not get exactly 100% speedup. Lets say the system produces 15 WUs each day, up from 8.

Great it seems, but consider that if you were to track the WUs individually, each would now take a bit more than 3 hours to complete. So the performance of the single thread goes down a little while the throughput of the system goes up by 87.5%. This can be scaled up all the way to 8 cores with 2 sockets. The more cores/threads you add, the more throughput you get, but the scaling is never 100% which means that the single threaded performance goes down a little each time as you load the system to it's maximum.

This is where the system architecture plays such an important role. Opterons scale better than the FSB Xeons because each Opteron has it's own memory controller, while the old Xeons all have to share the same memory controller in the chipset. So in the example above, a Xeon might do 9 WUs/d with a single thread running but only 14 WUs/d with 2 threads running FOR EXAMPLE. Nehalem Xeons have very simular system architecture (memory controller and interconnects) as the Opterons so they get the same or even better scaling in addition to having higher single-threaded performance compared to the old Xeons and especially the Opterons.

Now with regards to Hyperthreading. Lets say it gives you a 20% speedup. It's a fairly common, perhaps even upper bound for most software that can take advantage of it. But to take advantage of HT on our hypothetical system you need to run twice as many F@H clients because they only have 1 thread each. So we get this with our hypothetical machines:

1 socket, 1 core, 1 thread : 8 WUs/day. 1 WU takes 3 hours to complete
2 sockets, 2 cores, 2 threads: 15 WUs/day. 1 WU takes 3:12 hours to complete
2 sockets, 2 cores, 4 threads: 18 WUs/day. 1 WU takes 5:20 hours to complete


For F@H it doesn't really matter that each WU takes longer to complete as you add cores/threads because all you care about is getting the most WUs done every 24 hours. For a real-time single-threaded application like a gameserver it makes all the difference however. If the 2-core/thread 1 WU performance is the minimum needed then using HT on such a machine would only hurt you. You would need to use a much higher clocked CPU to make up for this shortfall, but on the plus side it would be able to handle twice as many gameservers so you would only need half the servers.

I dare not speculate however if this would work with your specific software. You may run into other bottlenecks with so many gameservers running on a single server.

Anyway, I hope this explains.

 

[DrMrLordX]

Running 8 single-threaded applications on an 8-core server is not a whole lot different than running 1 application that supports 8 threads.

Ding ding ding, we have a winner.

I was going to say the same thing, but once he figured out that the 2 ghz Gainestown system might be within his budget, I figured I'd keep my mouth shut and let him just buy that. Gainestown will have a huge advantage over Harpertown when pegging all cores with 8 threads just from the QPI advantage. Obviously Opterons will also start looking better under the scenario he has proposed, but unless he can get a sweetheart of a deal on a good 2P Opteron system at 2.3 ghz+, I don't think it will stack up all that well vs. Gainestown.

I was also thinking like aigomorla in that enterprise gear might not be the way to go, but if he wants 8 cores in a 1u rack then I guess he's sort of stuck with it.

 

[Trizik]

very much appreciated, jones377

now i'm just debating between E5504 and E5520 because single-threaded performance is very important and i'm concerned 2.0GHz won't cut it when the 32 slot servers are full, but E5520 increases my spending by $310. are the added benefits of E5520 (+250MHz CPU clock, +4MB L3 cache, +1.06 QPI, +266MHz DDR3 clock, Hyper-Threading, Turbo-Boost) worth the $310 for single-threaded performance? i'm guessing yes

here's a comparison of the two models:

E5504
Speed: 2.00GHz
L3 cache: 4MB
QPI: 4.8GT/s
DDR3: 800MHz
Hyper-Threading: no
Turbo-Boost: no

E5520
Speed: 2.26GHz
L3 cache: 8MB
QPI: 5.86GT/s
DDR3: 1066MHz
Hyper-Threading: yes
Turbo-Boost: yes

 

[alyarb]

You specifically said hyperthreading was NOT good for your application. If this is still true, be prepared to turn it off.

 

[PCTC2]

As far as Turbo, you probably won't notice much with full load of all your large game servers and the plethora of smaller game servers. The cache and clock may help, but again, don't know about the load of game servers. Hyper-threading probably won't help you with the single threaded-ness of your applications under full load. It may dip below your standards of the performance per server.

I currently use a dual E5520 system (much more expensive than yours running many virtual machines) but I think either would fit your bill. Just save some money and pick up the dual E5504 system. Unless you think the extra cache will help you.

 

[Trizik]

thanks PCTC2

You specifically said hyperthreading was NOT good for your application. If this is still true, be prepared to turn it off.

right, hyper-threading isn't good for the app's performance. i was just pointing out the features that 5520 has that 5504 does not. if i go with 5520 i will most likely disable HT.

 

[alyarb]

well i really don't think the cache is worth all that extra money. Your E5504/E5506 will not be slower than harpertown even if harpertown has 12MB L2.

 

[Trizik]

even though the Harpertown is 2.33GHz while the E5504 is 2.0GHz?

 

[DrMrLordX]

even though the Harpertown is 2.33GHz while the E5504 is 2.0GHz?

Yes. You're talking about two separate microarchitectures and platforms, one of which yields better performance per clock while also providing much greater memory bandwidth to keep all the cores fed when they are all loaded (Nehalem/Gainestown). There's also the issue of cache coherency, but Gainestown wins there as well.

The Front Side Bus represented the one major liability left in Intel's enterprise (workstation/server) systems. QPI fixed that.

To put it in perspective, if you were to compare a 2P, 8-core Harpertown system at 2.33 ghz to a 2P, 8-core (no HT/HT deactivated) Gainestown system at 2.0 ghz running just one of your game servers, it might be neck-and-neck, or maybe the Harpertown would have a minor edge. Once you load up 8 servers (one per core), you start saturating the FSB on the Harpertown system while trying to feed all 8 cores, slowing the performance-per-core on the Harpertown box. On the Gainestown system, it should handle the workload of 8 game servers in stride, giving it a significant performance advantage over Harpertown in that scenario.

 

[aigomorla]

To put it in perspective, if you were to compare a 2P, 8-core Harpertown system at 2.33 ghz to a 2P, 8-core (no HT/HT deactivated) Gainestown system at 2.0 ghz running just one of your game servers, it might be neck-and-neck, or maybe the Harpertown would have a minor edge. Once you load up 8 servers (one per core), you start saturating the FSB on the Harpertown system while trying to feed all 8 cores, slowing the performance-per-core on the Harpertown box. On the Gainestown system, it should handle the workload of 8 game servers in stride, giving it a significant performance advantage over Harpertown in that scenario.

LIES!

each time i tell the AMD people that, i get flamed... :\

(sacrasm.... now if u can tell me how i can explain this to the amd users so they understand...)

Hey OP:
1U rack is HELLLLLLLLLLLLLLLA LOUD.

I hope this box isnt going to be in your room, or have fun listening to airplane turbine engines...

And no im not kidding...

 

[DrMrLordX]

LIES!

each time i tell the AMD people that, i get flamed... :\

(sacrasm.... now if u can tell me how i can explain this to the amd users so they understand...)

Depends, which AMD people? I consider myself "AMD people" in that I consistently buy AMD machines, but even I have to admit that QPI represents a huge leap forward for Intel's enterprise platforms.

If they're willing to listen (and not all of them are), I'd say that QPI does for Intel's platforms what HT did for AMD's. They may cry foul over that (and HT3.0 does have some features, like HT Assist, that Intel either does not have/doesn't yet have/does have but doesn't do as well), but in the end they're mostly splitting hairs. QPI equals more memory bandwidth, lower memory latency, "real" IMCs on the chips that matter, and faster/better inter-core communication as compared to older generations of Xeon platforms.

So basically, Athlon MP is to Opteron as FSB Xeon is to QPI Xeon. Or at least, that's my way of looking at it.

I would think it highly amusing that any AMD fanboi would ignore the advantages of QPI over FSB considering how many AMD proponents bashed Intel for their use of a FSB on enterprise platforms (and yes, many of them went after multiple FSB platforms as well).

 

[aigomorla]

lol... i was saying sarcasm.... :X

 

[DrMrLordX]

lol... i was saying sarcasm.... :X

I know *) But you wanted me to tell you how to explain it to AMD people so their heads wouldn't asplode, so I took a whack at it. Honestly I don't know if they like the idea of Intel's transition from FSB to QPI as being a similar improvement and AMD's leap from Athlon MPs to Opterons, but it seems similar enough to me.

 

[aigomorla]

OP to be honest with your thread.

I would grab a PHII system, throw it in a small box, and have that setup.
If you need another, get another PHII system throw that in another box.

Keep on repeating until you meet requirements.

You can litterally build a PHII system for about 500 dollars.
Even unlock tri cores and make them quadcores.

That i see as your best possible route for a game server. Unless u want to become Blizzard and host your own WOW.

 

[Dadofamunky]

LIES!

Hey OP:
1U rack is HELLLLLLLLLLLLLLLA LOUD.

I hope this box isnt going to be in your room, or have fun listening to airplane turbine engines...

And no im not kidding...

He's right, the skinny rectangular PSes for 1U chassis are ridiculous, and there's no getting around it.

Funny thing, I went by Santa Rosa Computers today (in my local area) to buy a case and they have a stack of new Dual Opteron 2346 CPU 1U chassis with 2 GB ECC DDR 333 RAM and SATA ports for $150 each. Yes, the CPUs are installed. I know this isn't what OP is looking for but strangely enough it's kinda in the ballpark. I have to admit, those things are mighty tempting.

A is right, the PHII option is definitely the most cost-effective. 1U isn't happening there.

 

[jones377]

OP to be honest with your thread.

I would grab a PHII system, throw it in a small box, and have that setup.
If you need another, get another PHII system throw that in another box.

Keep on repeating until you meet requirements.

You can litterally build a PHII system for about 500 dollars.
Even unlock tri cores and make them quadcores.

That i see as your best possible route for a game server. Unless u want to become Blizzard and host your own WOW.

I got the impression that he was starting a game hosting business and was planning to populate an entire cabinet with 1U 2-socket servers. Different rules apply in that case.

 

[aigomorla]

I got the impression that he was starting a game hosting business and was planning to populate an entire cabinet with 1U 2-socket servers. Different rules apply in that case.

which is why i said if ur gonna be like blizzard and host your own WOW, then he will need a full blown gainestown.

However no matter how powerful the server is... you cant get past latency in your network connection.

So if you have a crap network, no godly server will save you.

 

[Trizik]

i'm trying to keep my monthly costs down and it's cheaper to colo 1 x 1U than 2 x 1U, 1 x 2U, etc.

pricewise a single 1U system for under $1,500 is going to work out better for me in the long-run, especially because shipping is expensive from my location

 

[alyarb]

ok, but i take it this is your first 1U machine? it'll run loud. where are you going to put it?

 

[aigomorla]

OP 1U is insanely loud...

Your underestimating your noise factor.

You will hear it in the room next door with the door closed.

 

[Trizik]

yep, first 1U machine

it's going in a rack @ colo