PrimeGrid
Please visit donation page to help the project cover running costs for this month

Toggle Menu

Join PrimeGrid

Returning Participants

Community

Leader Boards

Results

Other

drummers-lowrise
1) Message boards : Number crunching : World Expo Challenge (Message 85427)
Posted 1484 days ago by Profile Wave Fusion
what does cleanup work actually mean in regards to scoring positions?
Is it to essentially confirm the results currently on the leaderboard? (as in, credits are removed in a task is found invalid)

Or is there potential for rank gain from these tasks?
If so, is that ONLY for people who finished the task within the 8 day challenge window
Or is there incentive for others to intentionally seek out cleanup work post challenge?
2) Message boards : Number crunching : World Expo Challenge (Message 85313)
Posted 1486 days ago by Profile Wave Fusion
It appears I will miss a GPU deadline by 10 to 15 minutes. I'm unable to overclock an R9 265X mobile card with any software I'm familiar with. If it existed though, the same die is used in much faster GPUs. I know it would be stable. But.
I'm pretty disappointed after so much careful planning, I will miss a task so close to finishing.
3) Message boards : Number crunching : World Expo Challenge (Message 85270)
Posted 1489 days ago by Profile Wave Fusion
Well friends, the point of which CPU tasks newly started can no longer be completed in time for challenge points fast approaches. For most of my machines, that time has arrived.

Lets say your current crop of tasks is set to finish 20 hours before the deadline, and a new batch would take 40 hours.
Is there anything, any strategy, that would enable you to use those 20 hours for the challenge? =)
4) Message boards : Number crunching : World Expo Challenge (Message 85234)
Posted 1490 days ago by Profile Wave Fusion
Nice idea: This strategy will test CPU/RAM memory contention and find out if more tasks at lower speeds will bottleneck. Genefer heats up DRAM more than LLR by a few C degrees.

CPU TDP accounts for the IMC/DRAM. Keep an eye on IMC/bank/DRAM temps and power usage - Genefer compute with C7 2133 dual channel (2 modules) IMC requires nearly 35W on a 3.0GHz quad Haswell. (60-130W for CPU/DRAM/IMC) A dual core Ivy Bridge IMC at 2.8GHz with C11 1600 is around 15W. (25-55W CPU/DRAM/IMC)

Lower CAS RAM shorten up Genefer runtimes. The higher the CAS (and slower RAM speeds) the longer the runtime. Larger LLR tasks also see faster runtimes when RAM is rated for higher speeds with a lower CAS.


I'm unable to be certain the bottleneck was the RAM, but I think its most likely the cache on the CPU itself (L1, L2, and L3) The reason for this is because my laptops all either use 1333MHz at 9 CAS or 1600MHz at 11 CAS. The RAM many desktop users have is vastly superior, but seems to bottleneck at roughly the same number of workunits.

By lowering the clock speed, the CPU can use a much lower core voltage, the machine I'm writing from is only at 0.747 V, and this is regardless of how many cores or workunits are active.

The fastest Genefer task at the higher clocks, finished in about 65 hours. The current crop appears set finish around that time, perhaps as long as 70 hours, but, by running twice as many of them, will still pay off nicely.
My other two computers, using Ivy and Sandy Bridge CPUs are about 10 hours slower, but appear similar.

All I've done is lower my core speed & voltages enough to use all 4 cores within a much smaller thermal allowance than a typical desktop would have.
Even though I can't run my CPU as hot as a desktop, the lower speed results in a reduced demand for RAM & cache resources, so on the whole, my CPU is better fed for some types of workunits.
This means while I'm not the fastest, I'm not completely left in the dust either.
5) Message boards : Number crunching : World Expo Challenge (Message 85201)
Posted 1491 days ago by Profile Wave Fusion
Now that my cache now only has three active total tasks: 2 for CPU and 1 for GPU. My OpenCL Genefer GPU runtime to complete is 5hr 55min whichs includes crunching two CPU. (rather than 8hr with all four CPU "active" or 6hr 55min with two crunching and 2 suspended tasks in cache.)

For the fastest GPU or CPU runtimes: check to see if any suspended tasks are out of CPU memory pool. A 2hr+ runtime difference can be avoided if BOINC cache has only active tasks and no suspended tasks that are active in the memory pool.


Hmm. I don't know if the CPU can be "tricked" like that. around a hundred processes are in memory at all times, but the CPU doesn't stumble over any of them. I'm not sure that's what caused your speed increase.

In any event, I'm trying a radically different strategy.
Instead of running 2 CPU WUs at highest clock speed I can, 2.45 GHz, I'm running 4 CPU + 1 GPU at 1.55 GHz and getting the exact same temps.

The theory going forward is that if slowing the clock speed 900 MHz allows me to double the tasks I'm running, it will outweigh the increased runtime.
It's also gambling somewhat that calculation speed may NOT scale in a linear fashion with clock speed
I believe I'm gaining more than I'm losing, but we shall see.
6) Message boards : Number crunching : World Expo Challenge (Message 85176)
Posted 1491 days ago by Profile Wave Fusion
A 970 GPU only is nearly 1.5hr faster compared to [2]CPU+GPU. GPU/MCU/BUS usage for GPU only is the same when two CPU instances are computing at the same time.

Running GPU only - the time difference is surprising. At the same core 1539MHz clock: an OpenCL Genefer running on GPU only will finish in 6hr 20min instead of 7hr 55min (the CPU computing two instances.)


That's actually very useful knowledge. It might be only 1.5 hours difference for you, but over a much longer run time, say, 24 hours instead of 6, running CPU tasks would mean an extra 6 hours, for a total run time of 30 hours.

It would be even more pronounced if you loaded all 4 cores instead of just the 2.
The question I have now is if there's a difference for faster (3 GHz and up) processors and slower ones (2 GHz and below)
A slower one might be less demanding on other resources
7) Message boards : Number crunching : World Expo Challenge (Message 85168)
Posted 1492 days ago by Profile Wave Fusion
I don't believe the CPU was at fault here. I think BOINC started the OpenCL application, but the GPU was simply unresponsive, and it continued to count up on the progress bar, but wasn't actually doing anything,

The would explain why the CPU could run its 4 instances and not get too hot. It shares a heat pipe with the GPU, so ordinarily it throttles down if I try to run too much at once.
It was detectable, but just unresponsive. I'd never run Genefer before, so I dismissed it as normal.

But in hindsight, I think it's just a bug from the switchable graphics multi GPU configuration. The system powered down the AMD GPU and was unable to restart it.

After crashing/restarting the computer though, it's running as expected.
The GPU is showing signs of life, and I'm only running 2 of the CPU version to keep temps in check
8) Message boards : Number crunching : World Expo Challenge (Message 85161)
Posted 1492 days ago by Profile Wave Fusion
Ok I figured it out.. sort of. The GPU workunit said it was running, and the progress bar acted as though it were. But I'm highly skeptical it really was. Unless I have a 3rd GPU I don't know about.

The R9 265X showed few obvious signs of being under load. Its clocks were at idle 300 MHz, 0% utilization. very low (but some) thermal output. and BOINC said the entire time over 30 hours its CPU time was 00:00:01 which from running other programs seemed odd

Also odd was the running the CPU at full tilt without any thermal issues. They share a heat pipe, so that wasn't right..

Well. I crashed my PC trying to run GPU-Z. But I had the good sense to pause the workunits first.
The CPU ones are fine, but the GPU one has restarted from zero
The behavior is completely different, the GPU is hot and I can't run more than 2 CPU tasks alongside the GPU without going over the 95 C thermal limit

It worries me I should check on the other machines now.. >_>
9) Message boards : Number crunching : World Expo Challenge (Message 85160)
Posted 1492 days ago by Profile Wave Fusion
This is the first time I've GPU'd Genefer: I see OpenCL relies on a full CPU core. I'm running two CPU instances for maximum speeds instead of pinning the CPU at 100% - interrupts and memory contention happen infrequently. 3 CPU instances and 1 GPU on a quad (non-HT) will often miss - leading to longer total/CPU times.



I'm not sure how precisely to detect a potential memory bottleneck, and I was skeptical, however.. I have a sample that is unmistakable in its results

On twin systems with an Intel i7-4710HQ at 2.45 GHz (Haswell) and an AMD R9 265X
The first computer has a GPU and 4 CPU workunits running at the same time
The second computer has only the GPU running.

The second computer finished its workunit after 26.7 hours. while my own primary computer is on 37 hours and still only 88% done
I can't think what would account for this besides a GPU potentially underclocking, or some sort of memory bottleneck on the first computer


Is there some way to actually see this?
I expected a full load to be slower on the CPU side, but I am concerned that it would affect the GPU workunit so drastically
10) Message boards : Number crunching : World Expo Challenge (Message 85132)
Posted 1492 days ago by Profile Wave Fusion
I think spec wise, a Tahiti card is a great all-rounder, but I really favor low wattage cards that are stuffed with large amounts of utility, like unusually good double precision. a 250W card is too much.

I'm sure process improvements in 28 nm means newly manufactured R9 280s -could- be more efficient, unlike the laptop world, process improvements appear to be tuned to higher performance versions of that GPU, which command an even higher premium.

My older laptops used to ship with unlocked clock speeds, so the GPUs could often be clocked nearly as high as their desktop cousins, and remain stable without overvolting, because the more flawless chips were better suited for mobile systems.
It seems now if I want pure unrestrained performance I need to pay extra. For every little thing


Next 10 posts
[Return to PrimeGrid main page]
DNS Powered by DNSEXIT.COM
Copyright © 2005 - 2019 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 1.49, 1.44, 1.44
Generated 26 May 2019 | 11:59:12 UTC