not sure if it is worth a new thread and since it is technical i post here

i have seen by chance that the CEP2 WUs seem to be using really much RAM bandwith, so maybe, i thought, there is not my cpu power limiting the calculations, but my memory speed.
I worked on some measurements of cpu performance counters on a i5 2400 with ubuntu and let a CEP phase 2 Workunit run to have some 'real world' load. I've seen that the cpu core was not saturated, less than 500ms per second via powertop, also the energy consumption measured via performance counters was less than assumed, but the system did feel like it is under heavy load. So there was some bottleneck somewhere. I found a very high usage of the L3 cache bandwith on the core that was running CEP, but not enough to be a bottleneck. Sadly i can not measure the memory bandwith directly, the counter or the reader of this countergroup seems to be broken, but after some memory benchmarks it seems to be cleary the memory bandwith that is saturated for one core.
One core could use a bandwith of about 15GB/s at sizes over 6MB which is a sign to RAM instead of the L3 cache. Two and more cores get about 17GB/s.
Since i measured 14 to 14,5 GB/s running CEP it is comprehensible to assume that memory bandwith is crucial for theese workunits isn't it?
Additional that would lead to that running one WU at a time is enough, since the bandwith is already nearly fully used and running more would only slow them down (and lead to a little bit cooler cpu...).

But, that is why i am posting, maybe you know more than theese speculations, maybe the linux version of the application is even acting differently than the windows one and so on.
At the moment i am trying to run CEP unlimited, so at some times three workunits at once and my system is pretty happy to lag here and there, but the crunchtimes seem to be ok, in opposition to what i would assume, i see 2-4 hours runtime with a little bias towards 3 hours. Theese times on my home pc under windows with an i5 3570k. I don't believe that the cpu or the little bit faster ram (DDR3-1600 instead of 1333er) is able to get rid of this maybe-bottleneck. only about 20% faster with three times the workload.
Maybe the developers have run benchmarks of the application, so you can tell if it is really memory bandwith limited or if i'm simply wrong about something.

I apologise if i made speaking/writing errors, i'm not a native speaker. I hope it's still readable and understandable.

edit: i forgot something, the WUs have some strange behavior in their shown progress, i've never seen some WU over 30%, but 26.x% and some moments later it is finished, but seems to be valid and ok.

Hello bigeagle, CEP2 is split into 16 modules. Each module computes some different data about the same molecule. At the end of each module a new check point is written. Since some modules can take several hours while most are much shorter, this means that a lot of time might be lost if it is necessary to begin from the last check point. Sometimes the program reaches the end of a module and seems to give up. Perhaps the module could not compute correctly. For whatever reason, when the program gives up, it writes blank files for all the later modules, then sends the result back to the project.

Lawrence

On the RAM question, i really meant bandwith, not the amount of used memory
it's confusing me, since 2 WUs at once doesn't seem to run faster than 3 or 4, which would lead to that i made a mistake somewhere

the progress is clear now, thanks

If you hibernate or sleep your system you never loose a second. If you don't, you get your experience, long intervals before progress is saved [one of the reasons why this project is explicit opt-in ;].

As some WU (work units) for CEP2 last 2 hours and others last 16 hours, it is inaccurate to judge by how much time as passed before progress is saved. Also the estimated time remaining for CEP2 is...odd. Don't trust that number. It seems the file completes when it is good and ready to complete. :)

As I don't know the answer to the reasoning, nor accurately check my system closely enough to know, I just hibernate my system. In the past, when my PC was in the bedroom, the desktop power button light during hibernation would go up and down stopping me from sleeping. If you have that issue toss a shirt of it and problem solved.