I have been checking over my stats here and i can see that my fast computers with Xeon Processors that processes 4 times more workunits than my "workstation" pc:s also gets 3-4 times less credit per workunit.. why is that? I thought they should get the same ammount of credit for the same work performed??

Se the stats below...

GenuineIntel
Intel(R) Xeon(TM) CPU 2.80GHz [Family 15 Model 2 Stepping 7]
( 4 logical cpu:s (processing 4 workunits at a time)
Linux
2.6.22-14-server

cpu time claimed granted credit
9,192.43 12.19 11.12



GenuineIntel
Intel(R) Xeon(TM) CPU 3.06GHz [x86 Family 15 Model 2 Stepping 5]
( 2 logical cpu:s)

Microsoft Windows Server 2003
Standard Server Edition, Service Pack 2, (05.02.3790.00)

cpu time claimed granted credit
9,168.33 24.30 11.18



GenuineIntel
Intel(R) Celeron(R) CPU 420 @ 1.60GHz [x86 Family 6 Model 22 Stepping 1]
(1 logical cpu)

Microsoft Windows XP
Professional Edition, Service Pack 2, (05.01.2600.00)

cpu time claimed granted credit
9,877.32 27.29 33.24

are the xeons using hyperthreading?

The Celeron is one of the few Celeron's that is actually a pretty good CPU as it's based on Conroe. It'll be a much more efficient cruncher than the P4 based Xeons so you'll get more per cycle from it...

---

are the xeons using hyperthreading?

The Celeron is one of the few Celeron's that is actually a pretty good CPU as it's based on Conroe. It'll be a much more efficient cruncher than the P4 based Xeons so you'll get more per cycle from it...

The Xeon has 2 logical processor per processor.. So i guess they have that? It will mean in my example below the Xeon 2.8Ghz has 2 physical processors in the machine but it crunches 4 rosetta tasks at the same time and completes them at the same speed as the Celeron machine. But the Xeon does 4 at the same time while the Celeron does 1 !! :)

R@H is affected by memory performance. So I would try doing 2 workunits without the virtual machine? RAM is also important, do both Xeon's have 2-slot DDR2? or 4-slot FB-DIMMs (even better)?

---

with Hyperthreading on the Xeon, the CPU time won't be increasing at wall-clock speed - it'll be running somewhere around half that, so although two jobs are running at once, they'll both be running at half the speed, I'd expect. You might get some benefit from HT over running a single copy, but I don't think it's much (as far as Rosetta is concerned anyway, as the CPU only has limited FPU resources and that's what Rosetta is heavily dependant upon).

Not sure how much you know about how Rosetta packages tasks up, but in case you don't, two jobs that run for, say, 8hrs on a fast machine and a slow machine won't get the same credit, because the fast machine will generate more models (decoys) within that 8hr task. Basically a computer will run as many decoys within a task until it can't fit any more decoys into the time limit, so two 8hr tasks aren't necessarily equal. Each decoy is credited the same, so the granted credit is the credit per decoy multiplied by the number of decoys reported.

Your Celeron will process more decoys within a task than your Xeon will, although your Xeon is doing four at once. Your computers are hidden so I can't be any more specific though...

HTH
Danny

---

As I have pointed out to you, your xeon seems to be slower than my Core 2 Duo @3.2GHz.
I think your Xeon have HT on, but that would not help at all, since it only have 2 actual cores anyways.
The celeron you have is using the "Core 2" architecture, therefore they are actually pretty good.

with Hyperthreading on the Xeon, the CPU time won't be increasing at wall-clock speed - it'll be running somewhere around half that, so although two jobs are running at once, they'll both be running at half the speed, I'd expect. You might get some benefit from HT over running a single copy, but I don't think it's much (as far as Rosetta is concerned anyway, as the CPU only has limited FPU resources and that's what Rosetta is heavily dependant upon).

Not sure how much you know about how Rosetta packages tasks up, but in case you don't, two jobs that run for, say, 8hrs on a fast machine and a slow machine won't get the same credit, because the fast machine will generate more models (decoys) within that 8hr task. Basically a computer will run as many decoys within a task until it can't fit any more decoys into the time limit, so two 8hr tasks aren't necessarily equal. Each decoy is credited the same, so the granted credit is the credit per decoy multiplied by the number of decoys reported.

Your Celeron will process more decoys within a task than your Xeon will, although your Xeon is doing four at once. Your computers are hidden so I can't be any more specific though...

HTH
Danny

The thing you are talking about makes sense though.. My Xeon machines makes 4 WU in 2h 20min aprox. My Celeron makes 1 WU in 2h 20 min aprox... So in my eyes the Xeon is 4 times faster and should produce 4 times more credit.. But for 4 WU with the xeon i maybe get 40 credits while i get around 30 for 1 WU with the Celeron and thats very strange. Other projects give me more credit for the Xeon machines. So maybe i have to change to other project on those machines then.. Or try to turn off hyperthreading and see if that does any difference...

---

I can make a change to show my machines. But one question. Will the hostname of the machines be showed?

I can make a change to show my machines. But one question. Will the hostname of the machines be showed?

no - you can only see the hostname and IP if you're logged in.

---

I can make a change to show my machines. But one question. Will the hostname of the machines be showed?

no - you can only see the hostname and IP if you're logged in.

Ok, i have changed the settings now :)

---

your Xeon scores do look low - I'd recommend limiting the number of running Rosetta processes to the number of physical CPUs per box (setting available under BOINC preferences). I would expect that hyperthreading would have a beneficial effect by letting other processes run along-side Rosetta more smoothly, but with the added advantage that two Rosetta processes aren't competing for cache.

HTH
Danny

---

your Xeon scores do look low - I'd recommend limiting the number of running Rosetta processes to the number of physical CPUs per box (setting available under BOINC preferences). I would expect that hyperthreading would have a beneficial effect by letting other processes run along-side Rosetta more smoothly, but with the added advantage that two Rosetta processes aren't competing for cache.

HTH
Danny

If i run 2 processes of rosetta without changing the bios settings of the machine i will only run with 50% cpu power.. I need to turn of hyperthreading in the bios in that case. I wonder how Ubuntu 7.1 Server would react to such setting?

For the other Xeon machines i can not change that since those servers are in a production enviroment... But it would be nice to test it and see if it makes any difference..

---

You don't need to turn off HT in the BIOS is you can't. The 50% you see doesn't matter, as the CPU is really running close to 100% when you have 2 instances running.
Danny[/quote]

If i run 2 processes of rosetta without changing the bios settings of the machine i will only run with 50% cpu power.. I need to turn of hyperthreading in the bios in that case. I wonder how Ubuntu 7.1 Server would react to such setting?

For the other Xeon machines i can not change that since those servers are in a production enviroment... But it would be nice to test it and see if it makes any difference..[/quote]

yeah - as chungenhung says, the 50% in TM is misleading as there is only one physical CPU. In simple terms Hyperthreading just slots in other processes in any empty cycles and shows those processes as running on a virtual CPU. No need to turn off HT in the BIOS - just try reducing the max number of Rosetta processes to equal the number of physical CPUs and see if the RAC improves...

---

ohh.. :) I learn new things every day.. Thx.. Will try this :) But that also means that WU from other project should be finnished faster if i only run 2 tasks at once instead of 4.. Right?

---

ohh.. :) I learn new things every day.. Thx.. Will try this :) But that also means that WU from other project should be finnished faster if i only run 2 tasks at once instead of 4.. Right?

yeah - should do. HT can improve throughput, but AFAIK not for FPU intensive tasks, and two copies will increase cache swapping, so there are benefits to only running one copy per CPU too. I think you'll see more benefit from one task per physical CPU, but the HT should allow other (non-BOINC) tasks to run on the virtual CPU which will mean they interrupt BOINC less...

---

I have to agree about HT. Distributed computing uses 100% of each core. So when you do HT to have each core do two WUs, then they are going to be competing for CPU memory cache. There's no magic thing where a program that uses 100% of the CPU will let you run two programs and use 200% of the CPU.

A simplistic way to view HT and VM is that they add efficiency when non-intensive programs run on them. The big idea behind VM is if you have an office with 4 people on the internet and in word processors, you could run them all on one CPU instead of 4 CPUs.

---

I have to agree about HT. Distributed computing uses 100% of each core. So when you do HT to have each core do two WUs, then they are going to be competing for CPU memory cache. There's no magic thing where a program that uses 100% of the CPU will let you run two programs and use 200% of the CPU.

A simplistic way to view HT and VM is that they add efficiency when non-intensive programs run on them. The big idea behind VM is if you have an office with 4 people on the internet and in word processors, you could run them all on one CPU instead of 4 CPUs.

I thought the main idea behind VM, is to be able to run multiple OSs on the same hardware and at the same time.

Yeah you can do that too with VM.

---