Posts by Ant Evans

1) Message boards : Questions and problems : Profile recommendations for HPC (Message 101028)
Posted 9 Oct 2020 by Ant Evans
Post:
Run BOINC tasks at 100% as suggested and trust the scheduler. Maybe let users use something a little higher (lower, I guess, in Linux) so that BOINC tasks are pushed into the background. It works.

One other thing, think in physical cores, not threads, so you don't allocate resources you don't have. You will get better results if you stress all threads, but you will also increase the clock time to completion (see below) of any given task.

On the other hand, memory is less ephemeral and will be handled less elegantly. You want to have tasks complete in the shortest clock time possible to reduce memory thrashing and contention. Ways to do this: reduce the number of tasks, or radically increase the time between switching. Only uncheck 'leave non-GPU tasks in memory' if you have done the latter.

This all requires familiarity with boinccmd or a remote connection for the manager. No esoteric settings are needed.
2) Message boards : BOINC Manager : Feature: usage limits for when computer is in use (Message 99278)
Posted 13 Jun 2020 by Ant Evans
Post:
Turns out my own motherboard can police a max CPU temperature. Imagine my surprise.

It's an Asus B350 (AMD AM4) from 2018ish, updated a few times. (The option is buried three layers deep in the BIOS settings. Along with a few other things like ECC, it probably wasn't there when I bought the board.)

A BIOS max CPU temp is brilliant for BOINC. I set the max CPU temp to 50 degrees, no software needed. The machine throttles itself by changing the clock multiplier.

Make sure your cooling is up to snuff or you may get radical reductions in speed, way below what the power drivers usually alllow. For the same reason, I wouldn't choose a silly low number.
3) Message boards : BOINC Manager : Feature: usage limits for when computer is in use (Message 98423)
Posted 10 May 2020 by Ant Evans
Post:
... on the other hand, to add simplicity, if you could scale back core count when a user is active rather than having a Boolean choice as is the case today, you could get rid of the 'Suspend when computer is in use' checkbox, at least in the Manager. That's because you could limit the number of cores conditionally to zero.

I think that corresponds to <run_if_user_active>1</run_if_user_active>.
4) Message boards : BOINC Manager : Feature: usage limits for when computer is in use (Message 98336)
Posted 5 May 2020 by Ant Evans
Post:
For temperature you want TThrottle, unless you run an AMD CPU. It hasn't been maintained for a few years but it is still excellent. Trying to accommodate the various non-standards involved in temperature reporting is not something you would want to wish on the BOINC team.

Apparently some motherboards can do that.

I should have mentioned that even with the non-zero computer-in-use core count I described, you would still want the 'suspend when non-BOINC CPU usage is above X' setting.
5) Message boards : BOINC Manager : Feature: usage limits for when computer is in use (Message 98327)
Posted 4 May 2020 by Ant Evans
Post:
The BOINC client has settings to limit the number of cores and the CPU time. It allows for the suspension of computation while the computer is in use.

There is an implied asssumption that users who want to reduce resources available to BOINC while they're busy on the machine want to suspend BOINC completely. I have found that assumption not to hold in various use cases. I want to slow BOINC down, not stop it. But I don't want to slow it down all the time. I also don't want to fiddle with settings in normal use.

My requirement is to be able to set the number of cores available to BOINC according to whether the computer has a user active. Others may care more about a CPU time limit.

I have done this in the past by scripting, involving copying xml files and using boinncmd to update the client, all triggered by OS events. I don't recommend this.

So the suggestion is to bifurcate the available settings, creating two new ones.

global_prefs.xml
to <cpu_usage_limit> add <cpu_usage_limit_if_user_active>
to <max_ncpus_pct> add <max_ncpus_pct_if_user_active>

The 'user active' system trigger/state seems to work and wouldn't need to change. In the Manager, you'd want to split the Computing Preferences -> Usage Limits boxes vertically to add the new settings.

Apologies if it's already under development or has been discussed to death. All I could find was https://boinc.berkeley.edu/forum_thread.php?id=11471&postid=75964
6) Message boards : News : Rosetta@home studies coronavirus (Message 98326)
Posted 4 May 2020 by Ant Evans
Post:
I'm running Rosetta. It likes new hardware. It doesn't like old hardware.

The BOINC project doesn't necessarily agree that you have to roll your own.
https://boinc.berkeley.edu/trac/wiki/CloudServer

I've run BOINC in three clouds but I would recommend it mainly as a learning experience.

The BOINC client provides control best described as coarse. Modern OSs make up for most of that.
7) Message boards : Questions and problems : GPU and CPU Temperatures. (Message 93284)
Posted 23 Oct 2019 by Ant Evans
Post:
The maximum is not the optimum.

CPU: Consumer CPUs are not optimized for constant loads, and running BOINC at turbo frequencies 24x7 is inefficient. Underclock and undervolt where possible. If you are running Windows, it is trivially easy to restrict CPU frequency using a power profile, which will make the biggest immediate difference. You can easily automate switching profiles so that you have full speed when you are using the box for something else.

GPU: the same applies - underclock and undervolt where possible. How you do this will be vendor-specific. With less contended RAM you may be able to run a comparatively higher compute frequency than with the CPU, but the physics of voltage and power is even more brutal because of the larger die area.

Faster RAM is often the best performance return on electrons employed. The architecture makes a bigger difference than increments in speed. With a fat L3 cache or more memory channels, or a generation jump (DDR3 to 4), you will get a better marginal return on more compute cycles. But with consumer gear, faster compute is often just waiting frantically for the RAM to hurry up, sucking power in the process.
8) Message boards : Questions and problems : use all CPU power on 1 task at a time (Message 93282)
Posted 23 Oct 2019 by Ant Evans
Post:
LHC and Cosmology have multi-threaded executables. They run under VirtualBox only. Unfortunately my experience is that the VMs are badly behaved with more than one heavy thread, with or without hyperthreading. I had to limit tasks to one thread (LHC) or run them on an unattended box where user responsiveness doesn't matter.
9) Message boards : BOINC client : High boinc.exe CPU utilisation (Message 3459)
Posted 12 Mar 2006 by Ant Evans
Post:
I have had this problem too with 5.2.13 (Win32). I had boinc.exe using up to 2 hours of cpu time a day, with the same page fault and mystery I/O signs, and over 1GB of nominal free RAM, on a stable system. It was running 2 processes, 1 per CPU.

The only workaround seems to be to limit boinc to one process(or). If you only have one processor, limiting it to zero should do the trick too...

You can always put the other processor(s) to work on other projects. You don't want them sitting back and relaxing and getting ideas.

I started a post on the Manager thread about this problem, but the question belongs here. Original thread at

http://boinc.berkeley.edu/dev/forum_thread.php?id=630
_
10) Message boards : BOINC Manager : CPU utilization of manager 5.2.13 Win32 (Message 3307)
Posted 3 Mar 2006 by Ant Evans
Post:
Was it by chance doing a benchmark run at that time? As that's the only time that BOINC itself will take up CPU time.


No, that uses 100% of both processors at normal priority and your machine turns to treacle.

boinc.exe is now back to using 7-9%. I'll upload a screen grab of taskman when I can be bothered.

Obviously, I can't tell if this is the core or the gui. I can confirm that it is independent of the currently running work unit.
__
11) Message boards : BOINC Manager : CPU utilization of manager 5.2.13 Win32 (Message 3299)
Posted 2 Mar 2006 by Ant Evans
Post:
Damn, as soon as I posted that thread, it went back down to zero.

I guess as soon as I report the problem resolved, it will go back up again.

__
12) Message boards : BOINC Manager : CPU utilization of manager 5.2.13 Win32 (Message 3298)
Posted 2 Mar 2006 by Ant Evans
Post:
Since installing 5.2.13 I've noticed that boinc.exe is eating around 6-9% of CPU resources, as reported by taskman, against a previous 0% for v4.x. The extra load shows up as kernel time.

Since this is a dual CPU box, that's the equivalent of up to 18% of one processor, which is a lot to be stealing from projects.

I'm running 5 projects - the usual suspects. There are no bad work units, no transfers waiting, no suspect messages. I have noticed that LHC is spamming the server once a minute (now backing off), but that shouldn't tax the CPU.

What's yours showing? Is 5.1.13 a dog or is it just having a bad day?

Dual AMD Athlon MP with 1.5 GB, XP Pro SP2, DSL, blah blah.
__




Copyright © 2021 University of California. Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.