arrizza, you could test your 'improbable' theory that running Boinc in a Virtual environment, as well as normally increases performance:

Select to crunch HCC only on both the Virtual and local machine.
For one day run @ 100% on the Local Machine only.
Then run at 100% on both the Virtual Machine and the local machine for another day.
Compare the results (tasks completed, credit, run time).

Excellent reply, thanks JmBoulier!

This clarifies many things. I did notice the BOINC runtime didn't match up with reality, but I hadn't done the calculation to see how far the mismatch was. Thanks for clarifying.

In any case, the reason I'm brought this up is that there is an impact to my electric bill for running BOINC and if I can replace three machines with one doing the same "work", then I'm ok with that! And if I can simply add a VM or two and get more "work" out of the same physical machine then hey all the better.

I'm well aware of how multiple tasks operate in a time-sliced CPU and across multiple cores as well -- at least in the general case. An accurate analysis however does require some in-depth knowledge of the run-time characteristics of the particular applications on the particular machine and OS scheduling characteristics. I was hoping BOINC had a run-time pattern amenable to squeezing a bit more meaningful work out of the physical CPUs. I currently have 10 machines crunching (Cancer), all with various CPU capabilities. Almost all of these are all old machines (win95 anyone?!) that have since been retired and instead of throwing them out, I'm reusing them to crunch. I had created a VM for other reasons on one of the bigger machines with 6 cores and saw what initially looked like a net gain from having the VM crunch on the same physical machine. And you all have clarified, yet again, there is No Free Lunch (NFL).

In any case the question for me now is how do I measure the efficiency of a given machine so I know which ones to keep and which to replace?

From JmBoullier the correct measure is: tasks completed / calendar day

(yes: averaged over a period of time, keeping other run variants ceteris paribus, etc.)

1) Does BOINC report "tasks completed" anywhere? I've looked but can't find it.

2) Assuming that "tasks completed" is not reported:

It's clear I don't need absolute accuracy in the efficiency ranking of my machines but I do need relative accuracy between them.

So, even though the Results value is not a precise and accurate measure, is it close enough that I could still use it for this purpose?

In other words, if I have three machines A,B,C, each reporting 10 results per day, and one machine D reporting 30 results per day, would it be reasonable to draw the conclusion I can retire A,B,C and replace them with another machine of the same type as D? (hopefully dropping my electric bill in the process).

John

Thanks!


Yup, I've participated in one of them and actually made the same points you did.

But my question here is different. It's pretty clear that replacing a single core machine with a multi-core is better. But the practical question I need to answer is: should I replace 2 single-core machines or 3 single-core machines with a single six-core?

The only way to answer that is to have some objective measure of the efficiency of the individual single-cores vs the multi-core. Again, it doesn't have to be accurate in an absolute sense, just accurate in a relative sense, i.e. relative between machines.

I couldn't find a "tasks completed" value reported, so I'd have to somehow keep track of the tasks and then determine how many finished in a day. Not sure how to do that.

And so I was wondering if the Results value could be used as a proxy. In other words, do Results have a linear relationship with "tasks completed"? If Results, more or less, correlates linearly with tasks completed, then I can use them instead.

John

There's 2 ways to do this.

One method is to use the job_log_www.worldcommunitygrid.org.txt located in your BOINC data-directory, this file includes among other things cpu-time and the name of all WCG-tasks a computer has returned without any errors. But, with the huge variations between different tasks for some of the sub-projects, this often isn't a very good method...

Another method is to use the info available at the web-pages, under "My Grid", "My statistics", "Device statistics"... This will open-up an overview of your active computers. Just select one of them, and you'll get the 14 last days this computer got any WCG-credits, this can be exctended upto 365 days. While the info does include "Results returned", or more accurately results credited, with the huge variations depending on WCG-project you're running, this isn't a good indicator. But, The info also includes how many points given this day, and the cpu-time for this work. So, a fairly good indicator about your various computers production, just use Excel or similar to do the calculations of "Points / cpu-hours", find the average, and you've found how many points per hour per cpu each computer gets. To accurately see how much more (or less) a 6-way system gets, compared to 3 single-core systems, just multiply the average you've gotten for the 6-way with 6 (#cpus), and you've got a fairly good comparison...


BTW, just looking on each computers BOINC-Manager, in Advanced View on the Projects-tab, after selecting WCG and selecting Properties, as long as you're running v6.6.xx or later you've get info about Credit for Host, both total credit for host but also the Average credit. While by no means very accurate, comparing the Average host-credit for each computer will show your fairly resent daily production...

Just for information: I operated several years long a double Xeon Quad at 2.33 GHz with Win XP Pro x64 SP3. This host earned around 2'200 points daily.
I operate currently two Phenom II x6 with Ubuntu 10.04 x64. Until end of July 2011, the first host earned a daily average of 3'600 points with 2.8 GHz; the second host earned over 4'000 points.
In comparison, a P4 HT (3.4 GHz) earned around 500 points with 2 threads, and a Centrino (Pentium Mobile) with 1.6 GHz earned around 280 points. The last both with Win 2000 SP4.
The power consumption of the P4 HT was pretty similar to the Phenom II x6 2.8 GHz.
As already mentioned by Ingleside, you can easily use the device statistic page on WCG for collecting "historical" values and for calculating some "long term" average. In all cases, I consider a 10 days period as a minimum for a significant result.
Cheers,
Yves