1) Message boards : Number crunching : Setting up a local Squid to work with LHC@home - Comments and Questions (Message 47547)
Posted 19 Nov 2022 by Saturn911
Post:
How do you monitor this slowdown?

I did a tail -f to see access.log in real time.
What I noticed is:
At the end of a ATLAS WU there are a lot of get requests to atlascern-frontier.openhtc.io.
While uploading the screen scrolls slower then normal.
And yes: This may have other reasons.

typical upload speed: 4 MB/s
time used: ~45 s

My time per WU: 3 minutes
For 120WUs a day this means 6h a day.

btw I hate software that takes full CPU doing about nothing.

Don't blame me. Switched back to direct upload.
So Squid only serves CVMFS and web browsing.
But surely I will have a look at next releases of Squid.
2) Message boards : Number crunching : Setting up a local Squid to work with LHC@home - Comments and Questions (Message 47533)
Posted 13 Nov 2022 by Saturn911
Post:
Now divide 8.49% by 69 and you get 0.12 %
This is the average singlecore CPU usage squid spends per task.

Thanks for that clarification.
30 concurrently running ATLAS tasks on my side.

Peak performance for uploads only matters if squid can't saturate the upload direction of the internet line any more. I never noticed this.

No way; my upload line is limited to 1Mb/s by long VDSL-lines.
What I see by system monitor is 100% of one core as long as an upload is in progress.
The main problem is, this slows down all other requests to Squid.

Guess some kind of misconfiguration, but haven't found it yet.
3) Message boards : Number crunching : Setting up a local Squid to work with LHC@home - Comments and Questions (Message 47527)
Posted 13 Nov 2022 by Saturn911
Post:
This make me think you have same troubles.
Here my stats with squid atlas-uploads:

Resource usage for squid:
UP Time: 6801.577 seconds
CPU Time: 1154.614 seconds
CPU Usage: 16.98%
CPU Usage, 5 minute avg: 0.14%
CPU Usage, 60 minute avg: 20.90%
Maximum Resident Size: 1064416 KB
Page faults with physical i/o: 5

With direct atlas-uploads, squid average CPU usage is lower than 0.5%:
Resource usage for squid:
UP Time: 3845.516 seconds
CPU Time: 5.017 seconds
CPU Usage: 0.13%
CPU Usage, 5 minute avg: 0.01%
CPU Usage, 60 minute avg: 0.14%
Maximum Resident Size: 182528 KB
Page faults with physical i/o: 0

While uploading in progress CPU Time rises >90%
4) Message boards : Number crunching : Setting up a local Squid to work with LHC@home - Comments and Questions (Message 47524)
Posted 13 Nov 2022 by Saturn911
Post:
Now I tried the workaround like suggested:
"client_request_buffer_max_size 512 MB"
Works so far, but while uploading squid 5.7 takes 100% of one logical core.
Is this common behavior?
5) Message boards : Number crunching : Setting up a local Squid to work with LHC@home - Comments and Questions (Message 47503)
Posted 6 Nov 2022 by Saturn911
Post:
What about my solution?

Set for "lhcathome-upload.cern.ch" the no proxy option in Boinc-Manager.

On my side this works like a charm for weeks now.
6) Message boards : ATLAS application : Atlas Native Transient HTTP Errors Uploading Resultfile (Message 47285)
Posted 23 Sep 2022 by Saturn911
Post:

In case it's the latter, is it version >4.x?
In that case please downgrade to the most recent 4.x.


What's wrong with squid 5.x?
Gave it a try, but no uploads to ATLAS.
Had to switch back to 4.x :-(
7) Message boards : Theory Application : Theory native fails with \"mountpoint for cgroup not found\" (Message 45423)
Posted 6 Oct 2021 by Saturn911
Post:
Hello out there
I'm on Manjaro (Arch) Linux.
Here, for newer kernels, cgroup has changed from V1 to V2.
This ends with "\"mountpoint for cgroup not found\" for Theory native
while Atlas native runs o.k.

So I have to set kernel parameter
"systemd.unified_cgroup_hierarchy=0"
to enable cgroup V1.

Is it impossible to update Theory native for running on both (cgroup V1 and V2)?
I'm not a programmer so kindly excuse this (maybe silly) question.
8) Message boards : Theory Application : New Version 263.70 (Message 35890)
Posted 14 Jul 2018 by Saturn911
Post:
Possibly your desktop environment. Full featured desktops like Gnome or KDE use considerable RAM and CPU time. Lightweight desktops like LXDE and XFCE use considerably less RAM and CPU. Which desktop do you use?

Boinc is installed as a system service here.
So I don't need to log in to desktop for running LHC@home.
Btw, it's XFCE on the 6G machines while the 4G host runs without desktop environment.

You may navigate to the project's preferences page and set "max # of CPUs" to 1.

Since I did this I had no more "VM Heartbeat file specified, but missing heartbeat." errors.
For me it looks like we need at least a processor with more than two treads to run the 2-processor-mt-tasks.

Now I run two single core mt-task at once and its o.k.

Thank you a lot for your help!
9) Message boards : Theory Application : New Version 263.70 (Message 35824)
Posted 8 Jul 2018 by Saturn911
Post:
The new app is a multicore app that uses a 2-core setup on your hosts.
This is not recommended for the hosts you mentioned as they have only 2 cores.

You may navigate to the project's preferences page and set "max # of CPUs" to 1.

First I will give this a try.

About memory
Two of the failing computers are equipped with 6GB of ram.
I think this should be enough for a 2core Theory task.
One of them has 4GB only. But the result is the same.

Is it possible, that some other tasks of the OS (Manjaro Linux here)
blocks one of the CPUs and impedes VB to work correctly?
10) Message boards : Theory Application : New Version 263.70 (Message 35812)
Posted 7 Jul 2018 by Saturn911
Post:
For me too.
Since the mc apps 263.60 and 263.70 I have the "VM Heartbeat file specified, but missing heartbeat." trouble.
On three older machines, they all are dual core. Intel and AMD.
Most of the tasks failed but not all of them.
The computers are
ID: 10392891
ID: 10318807
ID: 10395493
Here one of the results

https://lhcathome.cern.ch/lhcathome/result.php?resultid=199736060

Before the mc apps these three worked like charm.

Please help.



©2023 CERN