FormulaBOINC

[Markfw]

Thanks !

OK< you now have 148 cores working on NF@home ! F@H may steal a few cycles for the video cards, but not near as much as WCG was hogging.

 

[TennesseeTony]

.....And how do you see individual boxes ?? The link above is just totals. All of this is new to me, so bear with me as I learn.......

This is where I got my info from: EDIT: looks like you may have figured this out though while I was creating this masterpiece.

 

[TennesseeTony]

Well I'll be darned, there is such a thing as BOINCtasks. I thought you meant BOINCstats. lol I'll have to look into that. Looks very useful!

http://www.efmer.com/b/boinctasks

 

[Markfw]

Since I last posted, apparently I have done 4,000 more, 31,396, thats 4000 per hour:

Edit, and I read the stats right, the Ryzen takes about .2 sec cpu, and the Intel Xeon about .32

 

[crashtech]

Yay, @Markfw to the rescue!

On less significant but perhaps humorous note, I did find a place for the other C2Q system, on the back porch with a wireless card. It's nice and cool out there, running 35°C at 100% CPU

I guess that makes 82 threads, all in.

 

[StefanR5R]

To track one's own progress, the number of tasks completed is not a meaningful number anymore. The task runtimes are too different. (There were more uniform on the day before the start of the race.) But I think it still holds true that credits/task is still roughly proportional to runtime/task or CPU time/task.

What's good though is that all NF@H WUs have a quorum of 1. I.e. unlike several other projects, NF@H does not require that a WU is run independently by two users. Once you submitted a result, the NF@H server runs a quick check on its own which appears to be a matter of minutes at most. And NF@H's "your computers" web page reflects this promptly too, as far as I can tell.

 

[StefanR5R]

tasks/core/GHz/day

I remember this post but was confused as to what you were trying to say. I get it NOW, after studying it a bit, or I think I do.

The metric "tasks / (core * GHz * day)" is actually the less important one in this post. In the end, PPD is important. I only posted the number of tasks because all of the tasks that went through these three boxes at that day were fairly uniform.

But now the tasks have become rather inhomogeneous. (Some long tasks in the mix, many 2-seconds tasks popping up too.) IOW "number of tasks", as well as "number of tasks normalized to core count and clock speed" are useful numbers for day 0 (the bunkering day) on those three boxes, but not in general.

Is the main point that the points awarded DO increase based on faster turn around times/faster hardware?

The main point was to check relative machine performance. I.e. answer questions like "should I go through the hassle of plugging in a moldy Core 2 Quad".

But one interesting bit in this observation was that credits awarded were seemingly not exactly proportional to work done, but slightly favored the faster machine. But whether there really is such an effect or whether it was just coincidence and noise, cannot be determined from just those three machines at just this one day.

Great idea though Ken! Still might prove useful though if someone runs some theoretical times.

It definitely is a splendid idea and, for all I know, might just prove true. But better measure twice before you cut once.

 

[StefanR5R]

Circa a day's work:

Core 2 Duo, 2.33 GHz:

15 tasks (3.2 tasks per core and GHz and day)
1,870 points (125 points per task)
[edit: ~400 points per core and GHz and day]​

Phenom II X4, 2.5 GHz:

39 tasks (3.9 tasks per core and GHz and day)
5,076 points (130 points per task)
[edit: ~510 points per core and GHz and day]​

i7-4960X, 4.5 GHz, 11 threads used:

193 tasks (3.9 tasks per thread and GHz and day, 7.1 tasks/core/GHz/day)
28,639 points (148 points per task)
[edit: ~580 points per thread and GHz and day, ~1060 points/core/GHz/day]​

Updated after almost 2.5 days' work:

Core 2 Duo, 2.33 GHz:

1,889 points/day
405 points per core and GHz and day​

Phenom II X4, 2.5 GHz:

5,310 points/day
531 points per core and GHz and day​

i7-4960X, 4.5 GHz, 11 threads used:

28,913 points/day
583 points per thread and GHz and day, 1070 points/core/GHz/day​

Spoiler: more detailed data

Core 2 Duo, 2.33 GHz:

running for 2d8h30m = 2.35 d, 2 tasks in parallel
tasks (total; per day):

48; 20.4 /d​

run time (total; min per task ... max per task; avg per task; total per day):

366,386 s (4.24 d); 2.08 s ... 18,712 s (5.20 h); 7,633 s (2.12 h); 1.80 d/d​

points (total; min per task ... max per task; avg per task; total per day):

4,439; 0.00 ... 227; 92; 1,889 /d​

405 points per core and GHz and day​

Phenom II X4, 2.5 GHz:

running for 2d11h30m = 2.48 d, 4 tasks in parallel
tasks (total; per day):

113; 45.6 /d​

run time (total; min per task ... max per task; avg per task; total per day):

773,839 s (8.96 d); 0.48 s ... 13,095 s (3.64 h); 6,848 s (1.90 h); 3.61 d/d​

points (total; min per task ... max per task; avg per task; total per day):

13,164; 0.00 ... 223; 116; 5,310 /d​

531 points per core and GHz and day​

i7-4960X, 4.5 GHz, 11 threads used:

running for 2d10h30m = 2.44 d, 11 tasks in parallel
tasks (total; per day):

318; 130.5 /d​

run time (total; min per task ... max per task; avg per task; total per day):

2,032,096 s (23.52 d); 2.04 s ... 40,505 s (11.25 h); 6,390 s (1.78 h); 9.65 d/d​

points (total; min per task ... max per task; avg per task; total per day):

70,475; 0.00 ... 1,405; 222; 28,913 /d​

583 points per thread and GHz and day, 1070 points/core/GHz/day​

Conclusions:

• The observations after 2.5 days concur with those after 1 day, except that...
• ...tasks/day have increased, surely because more of those nonsensical 2-seconds-tasks came through.
• But points/day has remained the same. This is because the award for each individual task depends on a metric which more or less closely relates to run time of this task. And the 2-seconds-tasks give only 0.00...0.01 points, i.e., inflate task count without impact on total run time or on total points.
  
• Points/day scale with thread count and CPU clock.
  
• In addition, a more modern machine yields somewhat more points/day when normalized to thread count and CPU clock. I.e., better IPC translates into better PPD to some degree.

Of course this is just for these three machines; YMMV.

And whether or not running more worker threads concurrently than the CPU has hardware threads is influencing PPD cannot be concluded from my data. I am curious what PPD @Ken g6 gets from his oversubscribed CPU.

PS: The fact that my PPD over a period of 2.5 days is practically the same as over 1 day indicates that testing over maybe 1/2 day could already be meaningful, provided that no over-long WUs are present during testing.

 

[StefanR5R]

If there is any more proof needed that I should get a life, here it is.

Spoiler: relationship between credits for a WU, and run time of the WU

Core 2 Duo: credits per WU = 0.012 s^-1 * run time
Phenom II: credits per WU = 0.017 s^-1 * run time
i7-4960X: credits per WU = 0.035 s^-1 * run time

This looks to me like credits awarded for a task depend on the work done, not on the amount of processor wall-clock time expended.

Notes:

• Both graphs show the same, just in a different scale.
• CPU time is always just a little less than run time, so I based the plots on run time.

 

[Markfw]

Well, overnight my points jumped to 108,631. Pausing WCG made quite the difference.

 

[TennesseeTony]

Both of the other teams' forums are strangely quiet, OCN for more than 24 hours now... I think Mark has them worried, and they are conducting all business behind closed doors or something.

Seriously though, we seem to be in a locked-step with 1st and 3rd, all moving up at roughly the same pace now. No doubt there could be some bunkering going on somewhere in hopes of providing a bit of a surprise at the end, but I think the final outcome will be about what it is now. Which is good, OcUK was making me nervous with all that recruiting!

 

[Markfw]

now at 112,281 since my last post 33 minutes ago. WOW

and interesting,, My Ryzen 1800x that is at stock, and has nothing else running is barely edging out my E5-2683 that is running noting else. 8 core beating 14, but at 3700 instead of 2500.

 

[crashtech]

now at 112,281 since my last post 33 minutes ago. WOW

and interesting,, My Ryzen 1800x that is at stock, and has nothing else running is barely edging out my E5-2683 that is running noting else. 8 core beating 14, but at 3700 instead of 2500.

Interesting, that would seem to indicate that Ryzen has better IPC at this particular task, quite a bit better, on the order of ~18%!

 

[StefanR5R]

Interesting indeed. Some workloads have shown a better scaling with Ryzen's SMT compared to Hyperthreading of Intel's current offerings. This could be such a workload.

 

[Markfw]

Interesting, that would seem to indicate that Ryzen has better IPC at this particular task, quite a bit better, on the order of ~18%!

Except the E5 has 14 cores, and the Ryzen only 8

 

[TennesseeTony]

I don't mean to toss water on anyone's campfire, but if you have 100 hypothetical tasks at 99%, and look at your score, then one shouldn't be too surprised to see a huge jump a few minutes later when those 100 hypothetical tasks have reported in. Hourly progress in nearly meaningless, just like in F@H. Even daily progress is near useless for projects that have variable size tasks, such as this one.

Great to see that Ryzen is a strong performer though! Glad to have AMD giving some competition to Intel. I just think the data given is too sparse to draw any conclusions yet.

 

[GLeeM]

Except the E5 has 14 cores, and the Ryzen only 8

Your Ryzen 1800 gets 3.4% of seconds for points. Your WU that took 5,900 seconds to finish got 200 points.
5900 times 0.034 = 204

Your Xeon gets 2% of seconds for points. Your WU that took 8,082 seconds to finish got 160 points.
8082 times 0.02 = 161

Most BOINC projects give ~1%. I think WCG gives less?

 

[TennesseeTony]

Wow, if that holds true across all tasks, 28 threads=2016 points per hour, versus 1958 for 16 threads. The Ryzen is a real beast to be so close with so many fewer cores! Can't wait to see the server offerings!

EDIT: Percentage of total run time converted to points

Haswell-EX Xeon- .019862% 2.5GHz
Westmere-EX- .022724% 2.93GHz
Haswell-E i7- .026741% 3.3GHz

 

[crashtech]

Except the E5 has 14 cores, and the Ryzen only 8

I tried to take that into account in my estimate. Clearly your Ryzen has a throughput advantage in this task even when compensating for core count and clock speed, how much is open to question.

 

[StefanR5R]

Broadwell-E @ 4.0 GHz gets 0.0367 points per second per thread. This means that Ryzen is exactly head to head with Broadwell-E in NF@H when normalized clock to clock: 0.0340 * 4.0/3.7 = 0.0367. (If the 1800X is indeed doing these tasks at 3.7 GHz.)

As noted above, Ivy Bridge-E @ 4.5 GHz gets 0.0347 points per second per thread. I.e. it gives the same points/thread as Ryzen 7 1800 X, but needs to be clocked higher than Ryzen to do so, and offers fewer threads.

On the other hand, something isn't right with this Ryzen 7 1700X host:
https://numberfields.asu.edu/NumberFields/show_host_detail.php?hostid=37380

• The host page states 1000 million ops/s float and int. At the 1800X host page, 4894 and 16267 million ops/s float and int are shown.
• The list of valid tasks of the 1700X show only 0.0069 points per second per thread.

 

[TennesseeTony]

Yep, Mark, that 1700X isn't being credited properly for some reason, only 1/4th what you should be getting. I hate to say it, but you may want to put it back to WCG. You can either remove the project or reset the project to get rid of all work (after "no new tasks" is chosen), this is under the "projects" tab of the advance BOINC Manager screen. I wouldn't even bother finishing up what it's currently running.

 

[crashtech]

Could it be that the info in that screenshot is from when WCG was running concurrently with Numberfields?

 

[StefanR5R]

Yep, Mark, that 1700X isn't being credited properly for some reason, only 1/4th what you should be getting. I hate to say it, but you may want to put it back to WCG. You can either remove the project or reset the project to get rid of all work (after "no new tasks" is chosen), this is under the "projects" tab of the advance BOINC Manager screen. I wouldn't even bother finishing up what it's currently running.

It is friendlier to the project to "abort" tasks rather than to than to "reset" the project:

• In the "abort" case, the project server gets notified that the client will not finish these tasks, and the server can then send these tasks to others right away.
• In the "reset" case, the project server is not notified. Hence the server needlessly waits until the WU times out.

So, first "abort" any WUs that you don't want to finish anymore, and then either "reset" or "remove" the project. (Could be though that "remove" automatically "aborts", I don't know.)

Could it be that the info in that screenshot is from when WCG was running concurrently with Numberfields?

Good question. I arrived at the number of 0.0069 points per second per thread by looking at NF@H WUs which were finished just shortly before I posted.

 

[4thKor]

You guys are definitely closing the gap!
I'll have to see if I can get some MOARRR POWAAAAA!

 

[StefanR5R]

Re Ryzen 1700X:
Could it be that NF@H miscalculates credits if the client hasn't been benchmarked?

I.e. View -> Advanced View -> Tools -> Run CPU benchmarks. This will pause all current tasks and do some brief benchmarking. When this is over and normal crunching is resumed, then I guess an "update" to the project is in order. Then check the host's web site again if more plausible float and int ops/sec have been filled in.

Normally this benchmarking should be done automatically by the client when needed.