are there any other threads running that could be causing rosetta to get swapped out of the cache? Anti-virus maybe (Norton has a reputation of being a bit of a resource hog)...

Not that I'm aware of. The system has 2GB of RAM and is only running R@H. I've got AVG installed but that doesn't show up as using any of the CPU time. The only way I check R@H is by using remote desktop so could it be not having a monitor plugged in?

Also, when looking at my results I'm getting low credit because the CPU isn't performing as many decoys as it says it should be. I'll probably install Vista 64 on it within the next few days and see if that has any impact.

Is there any way of benchmarking my systems performance and then getting some comparable results? What if I used SANDRA to take some benchmark scores and see if there is any throttling going on?

---

are there any other threads running that could be causing rosetta to get swapped out of the cache? Anti-virus maybe (Norton has a reputation of being a bit of a resource hog)...

Not that I'm aware of. The system has 2GB of RAM and is only running R@H. I've got AVG installed but that doesn't show up as using any of the CPU time. The only way I check R@H is by using remote desktop so could it be not having a monitor plugged in?

Also, when looking at my results I'm getting low credit because the CPU isn't performing as many decoys as it says it should be. I'll probably install Vista 64 on it within the next few days and see if that has any impact.

Is there any way of benchmarking my systems performance and then getting some comparable results? What if I used SANDRA to take some benchmark scores and see if there is any throttling going on?

not having a monitor isn't a problem, but is there a screensaver running in between RD sessions? You could try setting that to blank...

---

Is there any way of benchmarking my systems performance and then getting some comparable results? What if I used SANDRA to take some benchmark scores and see if there is any throttling going on?

i don't think a benchmark like Sandra would show throttling unless it's due to one of the cores getting too hot as the benchmark will cause the CPU to run at full tilt? Could it be temperature related? Could you drop the voltage slightly?

CPU-Z might be better to check the core speeds...

---

hmmmm... maybe the memory speed is quite important then... unless cool n quiet is kicking in to some extent?

No memory speed has little to do with Boinc, I have already done that with my own 9850!

---

Hi. I am "the other guy"! :)

No overclocking here - stock 2.5Ghz machine (200MHz Bus x 12.5 multiplier).

So I'm running Vista with 8Gb and Boinc 6.4.5 compared to XP with 2Mb running Boinc 5.10.45 - those are the basic differences.

My RAM is DDR2 running at 400MHz - 4 sticks of 2Gb Nanya PC2-6400 (400MHz) 5-5-5-18 according to CPUZ.

My FPU speed is showing 20% lower (OCing probably), but my Integer speed is nearly 20% higher. Something in there? No idea, but it doesn't sound right, does it.

Aside from that, my machine is on 247 but I work away half the week so there's literally nothing else running on it most of the time. I also tweak my start-up programs with MSConfig and the help of Bleeping Computer and my Services with BlackViper - I've found that very useful.

I'll PM nightmareXX with any more detailed information he might need. If I was him I'd be confused too.

Ah hey, thanks for replying :D

If you've looked at my Phenom recently, it'll be showing the speeds at stock. I dropped the OC thinking that it might be affecting the results but it didn't make any difference.

The only continuous feature between all the PC's I've looked at and been able to get in touch with the owner is running PC6400 RAM. I've pinched some from another PC and put that in the 9850 so hopefully that'll make a difference.

As far as that systems load goes, it's running BOINC and nothing else. It's not used what-so-ever so each R@H gets 25% CPU.

I have a 9850 currently running Linux and Rosetta that also does nothing else. It seems to be doing units in about the same time as yours is. What size hard drive do you have? Could it be close to being full? I noticed you have a 1 week cache while I have a 1 day cache, neither should make any difference over the long run. Does your machine connect to Rosetta every day or just when it needs new work? I also upped the memory speed in my 9850 and it made no difference that I could tell. Apparently memory speed is not a way to speed up Boinc.

---

Well I've done a few units with the DDR800 RAM and it doesn't appear to have made any difference. I'm fast running out of ideas now and it's starting to really annoy me :(

I just need to ask: the PC you linked to at the top of the page doesn't show any of the changes you've indicated you made. It's still running the old BOINC for a start and when you reported that there was no change I noticed that no tasks (maybe one) had actually been reported back at the time.

Which computer are you actually referring to when you talk about the changes you made? ID 832882 or ID 1018472? If the latter (running 6.4.5) there's no outstanding tasks to run. I'm confused too. If the former, how come it's showing you haven't even updated Boinc?

Are there any other threads running that could be causing Rosetta to get swapped out of the cache? Anti-virus maybe (Norton has a reputation of being a bit of a resource hog)...

That is the reputation. Trouble is it's the opposite of the truth (unless you have a Pentium 2 or 3 and Norton 2005 or older). If you want an AV that's faster in general operation, smaller, less intrusive, quicker to boot and more comprehensive then Norton has been tested to be much better (AVG introduces more delay into booting by a factor 3, for example). Check anything from av-comparatives to simple mag reviews to confirm that. It's really tiresome to hear the same FUD as well as being bad general advice. Uninstall AVG and put Norton on is better advice.

Answered another way, I'm running Norton 360 and you're all trying to find a way to get the credits I manage on a stock machine. Do the math.

---

I have a 9850 currently running Linux and Rosetta that also does nothing else. It seems to be doing units in about the same time as yours is. What size hard drive do you have? Could it be close to being full? I noticed you have a 1 week cache while I have a 1 day cache, neither should make any difference over the long run. Does your machine connect to Rosetta every day or just when it needs new work? I also upped the memory speed in my 9850 and it made no difference that I could tell. Apparently memory speed is not a way to speed up Boinc.

HDD is only a 120GB but it's no where near full and BOINC has lots of space that it can use.

As you pointed out, cache size doesn't make a difference as we're both getting low credit. R@H can connect as often as it likes, I'm on a broadband connection via my router.

However I had a thought. My motherboard only has a 1x HT link. If my memory serves me correctly, this is the speed at which the CPU talks to the rest of the system or something to that effect. Sid on the other hand is using this motherboard which has a max HT link of 3x. Cross referencing this with AMD's website, the 9850 is an AM2+ CPU. If you're also using a HT 1x motherboard, this is the only reasonable explanation that I can come up with.

Yes Sid, that's right. I downloaded the newer client of BOINC and added myself as an entirely new client and only let it get a few work units. I let those run, saw that I wasn't getting any more credit and reverted back to my old client (2 separate install locations).

For the moment I've set BOINC to only use 2 of the 4 cores. If I get more points when using just 2 cores, it'd agree with my HT link theory I think. Anyone got some more straws I can grasp at? :P

EDIT
I've browsed the active computers a little more and found another AMD with appropriate RAC. I've PM'd the owner to see if I can get some more information about the system.

---

I have a 9850 currently running Linux and Rosetta that also does nothing else. It seems to be doing units in about the same time as yours is. What size hard drive do you have? Could it be close to being full? I noticed you have a 1 week cache while I have a 1 day cache, neither should make any difference over the long run. Does your machine connect to Rosetta every day or just when it needs new work? I also upped the memory speed in my 9850 and it made no difference that I could tell. Apparently memory speed is not a way to speed up Boinc.

HDD is only a 120GB but it's no where near full and BOINC has lots of space that it can use.

As you pointed out, cache size doesn't make a difference as we're both getting low credit. R@H can connect as often as it likes, I'm on a broadband connection via my router.

However I had a thought. My motherboard only has a 1x HT link. If my memory serves me correctly, this is the speed at which the CPU talks to the rest of the system or something to that effect. Sid on the other hand is using this motherboard which has a max HT link of 3x. Cross referencing this with AMD's website, the 9850 is an AM2+ CPU. If you're also using a HT 1x motherboard, this is the only reasonable explanation that I can come up with.

Yes Sid, that's right. I downloaded the newer client of BOINC and added myself as an entirely new client and only let it get a few work units. I let those run, saw that I wasn't getting any more credit and reverted back to my old client (2 separate install locations).

For the moment I've set BOINC to only use 2 of the 4 cores. If I get more points when using just 2 cores, it'd agree with my HT link theory I think. Anyone got some more straws I can grasp at? :P

EDIT
I've browsed the active computers a little more and found another AMD with appropriate RAC. I've PM'd the owner to see if I can get some more information about the system.

I think if you look you and I are crunching units faster than Sid. He is taking about 12 to 14 thousand seconds to do a unit, I looked at several of his pages, while you and I are both down around the nine to twelve thousand second range. I think this needs more research. I wonder if it is the OS, I am running Linux while you are running XP and Sid is running Vista.

The link you gave is to a dual core machine taking about forty thousand seconds to do one unit!

---

sorry dual post, too fast on the clicks

Firstly, what motherboard are you using so I can check on it's spec?

I've just completed my first unit using only 2 of my available 4 cores. Just as I feared, it got the full points for the unit as the HT link wasn't being hammered as much. I'll let it run for a bit longer to see the effect only using 2 cores has but it looks like a new motherboard is in order.

---

I think if you look you and I are crunching units faster than Sid. He is taking about 12 to 14 thousand seconds to do a unit, I looked at several of his pages, while you and I are both down around the nine to twelve thousand second range. I think this needs more research. I wonder if it is the OS, I am running Linux while you are running XP and Sid is running Vista.

The link you gave is to a dual core machine taking about forty thousand seconds to do one unit!

The run-time is a preference you can set - the tasks consist of decoys (models) and the task will keep running new models until roughly the run-time you set. You get credit per decoy though so the run-time doesn't affect credit per hour. In other words, shorter task run-time = fewer decoys per task but more tasks. Credit per CPU hour is the same either way. ;)

---

Firstly, what motherboard are you using so I can check on it's spec?

I've just completed my first unit using only 2 of my available 4 cores. Just as I feared, it got the full points for the unit as the HT link wasn't being hammered as much. I'll let it run for a bit longer to see the effect only using 2 cores has but it looks like a new motherboard is in order.

that'll be because there's more cache available for tasks in the two running cores.

I think it's more likely to either be that Sid's queue is shorter than yours so he's running different (more recent) tasks which may give higher credit, something to do with memory - has anyone tested changing the timings to see if that makes a difference? If the cache is saturated (which it seems it is if running two tasks improves things) then maybe it's the memory latency that's the problem?

or could Sid's mobo possibly be OC'ing automatically? Quite a few do by default but not to that extent...

---

that'll be because there's more cache available for tasks in the two running cores.

I think it's more likely to either be that Sid's queue is shorter than yours so he's running different (more recent) tasks which may give higher credit, something to do with memory - has anyone tested changing the timings to see if that makes a difference? If the cache is saturated (which it seems it is if running two tasks improves things) then maybe it's the memory latency that's the problem?

or could Sid's mobo possibly be OC'ing automatically? Quite a few do by default but not to that extent...

Fair point about the cache but I'm pretty sure that the HT link is the speed that the CPU has to the RAM. I replaced the 667MHz RAM I had in there to 800MHz stuff and it's made no difference. R@H must do a lot of RAM data transfer and therefore, a faster HT link should provide more decoys in a given time as more data can be transferred.

As for RAM timings, my 9850 is on 5-5-5-15-23 where as my 940 is 5-5-5-15-22 so I doubt the timings have much effect on the RAC.

Sadly I have no way of being able to check this without buying a new motherboard and I've exhausted pretty much everything else.

As for the OC'ing, I doubt it's doing it automatically. AMD boards don't have that sort of feature.

EDIT
Done some reading around HyperTransport and it isn't related to the memory performance. It's the link from the CPU to the rest of the system, minus the memory so it can't be that. What I did find out was that unganged mode RAM allows for more requests per second so I'll try unganging my modules and see if that works.

---

However I had a thought. My motherboard only has a 1x HT link. If my memory serves me correctly, this is the speed at which the CPU talks to the rest of the system or something to that effect. Sid on the other hand is using this motherboard which has a max HT link of 3x. Cross referencing this with AMD's website, the 9850 is an AM2+ CPU. If you're also using a HT 1x motherboard, this is the only reasonable explanation that I can come up with.

I don't know anything at all about Hypertransport, but the way you've written this out makes it look like some kind of multiplier rather than a version number. Looking at Wikipedia it does seem to run at different speeds, with your 1.0 maxing at 800MHz while mine (below) is running at 2GHz.

What this actually means, I have no idea, but you may well be onto something.

I think if you look you and I are crunching units faster than Sid. He is taking about 12 to 14 thousand seconds to do a unit, I looked at several of his pages, while you and I are both down around the nine to twelve thousand second range. I think this needs more research. I wonder if it is the OS, I am running Linux while you are running XP and Sid is running Vista.

The run-time is a preference you can set - the tasks consist of decoys (models) and the task will keep running new models until roughly the run-time you set. You get credit per decoy though so the run-time doesn't affect credit per hour. In other words, shorter task run-time = fewer decoys per task but more tasks. Credit per CPU hour is the same either way. ;)

Yes, I have set my runtime to 4 hours rather than the default 3 hours.

If I'm understanding correctly, Mikey is saying that I'm taking (about) 205 processing seconds for each credit claimed, while NightmareXX is only taking about 190 seconds, irrespective of how many tasks end up going through. I have no idea if this reflects the slight overclocking, but it's certainly not reflected in the granted credit.

I think it's more likely to either be that Sid's queue is shorter than yours so he's running different (more recent) tasks which may give higher credit, something to do with memory - has anyone tested changing the timings to see if that makes a difference? If the cache is saturated (which it seems it is if running two tasks improves things) then maybe it's the memory latency that's the problem?

Or could Sid's mobo possibly be OC'ing automatically? Quite a few do by default but not to that extent...

It may be that some tasks grant better credit, but it's very consistent so I'd doubt it.

I've run CPUZ and there's no OC'ing at all - everything comes up bang on what it should from a stock machine (+- the odd decimal point):

Core Speed: 2500MHz
Multiplier: 12.5
Bus Speed: 200MHz
HT Link: 2000MHz

Fair point about the cache but I'm pretty sure that the HT link is the speed that the CPU has to the RAM. I replaced the 667MHz RAM I had in there to 800MHz stuff and it's made no difference. R@H must do a lot of RAM data transfer and therefore, a faster HT link should provide more decoys in a given time as more data can be transferred.

As for RAM timings, my 9850 is on 5-5-5-15-23 where as my 940 is 5-5-5-15-22 so I doubt the timings have much effect on the RAC.

Again I'm unfamiliar with how to report the RAM timings, so I'll just copy what CPUZ says for me:

Memory Tab, Timings:
DRAM Frequency: 400MHz
FSB:DRAM: 1:2
CAS# Latency (CL): 5.0 clocks
RAS# to CAS# Delay (tRCD): 5 clocks
RAS# Precharge (tRP): 5 clocks
Cycle Time (tRAS): 18 clocks
Bank Cycle Time (tRC): 24 clocks
Command Rate (CR): 2T

SPD Tab (as above) 400MHz 5-5-5-18-23 1.8V

I mention this because it's different to your figures, not knowing if it's better or worse than yours or provides any extra clues.

---

so what are we left with? Is Vista 64 that much better at scheduling a multi-core CPU than XP 32?

---

i've just done the maths on the latest page of results from each machine and the difference has changed:

The previous averages were 12.8 (NightmareXX) and 16.5 (Sid) credits per hour and they're now 15.2 and 16.6 respectively. I'd say it's probably down to the size of the tasks you're running - a shorter queue will get newer tasks which might be modelling something that fits into cache better and so requires less memory access etc.

NMXX - I'd expect your RAC to start rising...

---

So what are we left with? Is Vista64 that much better at scheduling a multi-core CPU than XP32?

It might be, though I do believe an update of the Boinc Manager is in order. I was getting a very high error rate with 5.10.45 which 6.4.5 has resolved and while there's no sign of that in any of our machines (to the point of erroring out) there may still be some benefit in addition to removing one of the differences between us.

One other question I wanted to ask is about the credit system. Is it only dependent on the hardware or the HWOS combination? There may be a much simpler answer than we think!

---

So what are we left with? Is Vista64 that much better at scheduling a multi-core CPU than XP32?

It might be, though I do believe an update of the Boinc Manager is in order. I was getting a very high error rate with 5.10.45 which 6.4.5 has resolved and while there's no sign of that in any of our machines (to the point of erroring out) there may still be some benefit in addition to removing one of the differences between us.

One other question I wanted to ask is about the credit system. Is it only dependent on the hardware or the HWOS combination? There may be a much simpler answer than we think!

Well I'll test that tomorrow. I'll leave my XP installation intact and move the 6.4 client over to a Vista 64 install. I'll then leave it a few units to see if the score improves. If it still doesn't, I'm 99% sure that I've covered everything else that it can be and I'll order a HT 3.0 motherboard and hopefully, that'll sort it out.

---

i've just done the maths on the latest page of results from each machine and the difference has changed:

The previous averages were 12.8 (NightmareXX) and 16.5 (Sid) credits per hour and they're now 15.2 and 16.6 respectively. I'd say it's probably down to the size of the tasks you're running - a shorter queue will get newer tasks which might be modelling something that fits into cache better and so requires less memory access etc.

NMXX - I'd expect your RAC to start rising...

Guys, did you read the above?

I don't think there's any difference other than the tasks you're running due to queue length. It looks like the newer tasks are giving you both similar credit...

---

Keep in mind, as dcdc says, credit is not an absolute thing. It is based on how difficult all the other machines out there are finding it to produce models for a given task. Each task and in fact each model has different code paths that may be taken or not, just depends. This is why there is significant variability from one task to the next.

The first thing to do, is measure one day of activity, change absolutely nothing (except you'll have different tasks) and measure the next day and see if your figures are consistent with themselves. You can often see 10% or more variation just based on the type of task you happen to be crunching at the time.

---

Rosetta Moderator: Mod.Sense