BOINC Pentathlon 2021

[GOTSpectrum]

10 minutes ago, Shlouski said:

Sidock wants an invitation code to join, on a glance I didn't see anyone mention anything about this in the thread, so how do I get one?

I didn't get anything like this, did you try to create your account through the BOINC manager? 

[leadeater]

31 minutes ago, Shlouski said:

Sidock wants an invitation code to join, on a glance I didn't see anyone mention anything about this in the thread, so how do I get one?

Crunch_4Science

[leadeater]

21 minutes ago, GOTSpectrum said:

I didn't get anything like this, did you try to create your account through the BOINC manager? 

If you create using BOINC Manager I think you can't login to their website? Inv code is above and can also be found on the project site, they are just using it for really basic spam account creation prevention.

[Lightwreather]

Well, I hereby bestow upon you command of 4 threads on my 6700. Use them wisely

[Shlouski]

11 minutes ago, leadeater said:

Crunch_4Science

 

Thanks

 

33 minutes ago, GOTSpectrum said:

didn't get anything like this, did you try to create your account through the BOINC manager? 

 

Nah, I was out the house so I went straight to the website.

 

Cpu wise I have:

 

3700x

e5-2680 v2

3500x

I7 8700k

1600af

I7 7700

I7 4790k 

 

Are any of these cpu's not worth using?

 

 

 

 

 

[Lightwreather]

You know it's funny that they are using this age-old iconography for connecting, I like it

[GOTSpectrum]

 

Let's see if Maddison will do a RT or shout out of some kind.  

 

EDIT: Can you guys RT this, or tweet other members of LMG/FP.  Someone has gotta be willing to shout out the pent right? 

 

[leadeater]

5 minutes ago, Shlouski said:

Cpu wise I have:

 

3700x

e5-2680 v2

3500x

I7 8700k

1600af

I7 7700

I7 4790k 

 

Are any of these cpu's not worth using?

Nice, very nice. All worth using if you can, up to you how much you want to contribute.

[NetoriusNick]

On 4/27/2021 at 10:57 AM, leadeater said:

 Quick Start Guide:

 Place Holder for new Guide

 

BOINC Pentathlon 2021

Announcements || Stats || Daily Bulletin

 

Dates: May 5 00:00 UTC - 19 May 00:00 UTC

 

4th Year for the LTT Team

Last year the team gain to 6th overall and gain Bronze in Marathon.  Lets go higher this time!

 

About the Pentathlon

The Pentathlon is an annual event for several BOINC teams to compete across five disciplines (five different BOINC projects) to gain points and win medals.  Several teams, but only one team comes out as the winner.  The event last for over two weeks.  If you up to the challenge, come on and join the team in competing in the pentathlon and see if your hardware can handle the heat.

 

BOINC Info and Guides

Team Name: LinusTechTips_Team

 

 

Announced

 

Marathon - SiDock@Home

Type: CPU Work Units

https://www.sidock.si/sidock/

Start: 05 May 2020 00:00 UTC
End: 19 May 2020 00:00 UTC

 

 

Javelin Throw - TBA

Type: 

Place:

Place Holder for Link

Start: TBA
End: TBA

 

City Run - TBA

Type: 

Place: 

Place Holder for Link

Start: TBA
End: TBA

 

Cross Country - TBA

Type: Typically GPU

Place: 

Place Holder for Link

Start: TBA
End: TBA

 

Sprint - TBA

Type:

Place:

Place Holder for Link

Start: TBA
End: TBA

 

 

All start times are 00:00 UTC

 

The Disciplines

 

• Marathon (14 days)
  • TBA
• Sprint (3 days)
  • CPU or GPU project with a quorum of 1
• City Run (5 days)
  • CPU project
• Cross Country (5 days)
  • CPU or GPU project
• Javelin Throw (5 x 1 day, only each team's third best daily score counts)
  • CPU or GPU project with a quorum of 1

CPU or GPU project with a quorum of 1 

you man you can only pick gpu or cpu as a team but not both?

[leadeater]

5 minutes ago, NetoriusNick said:

CPU or GPU project with a quorum of 1 

you man you can only pick gpu or cpu as a team but not both?

Any person can run as many projects as they like, even on a single computer. You can configure how much resources each project gets through various different settings but very generally speaking there is no restrictions on what you can run.

 

For this event it's just a coordinated effort to pick what we think is best to run and how many of us and for how long. We do that in this thread as the event is ongoing,

 

There is also the if in doubt run SiDock@Home project which is CPU as that is the marathon event that runs over the entire length of it, so if you are unsure this one will always help.

[RollinLower]

this year i might have some more hardware to throw at the marathon!

[Windows7ge]

2 hours ago, leadeater said:

@Windows7ge

Progress update:

• 1x 96 cores
• 1x 72 cores
• 7x 32 cores
• 5x 12 core

Total: 452 cores (cores not threads)

Enable Hyperthreading/SMT get that run time performance boost.

[leadeater]

3 minutes ago, Windows7ge said:

Enable Hyperthreading/SMT get that run time performance boost.

Can't most of them are VMs and you actually get a bit of performance regression if you create VMs greater than the total actual number of cores.

[leadeater]

12 minutes ago, Windows7ge said:

Enable Hyperthreading/SMT get that run time performance boost.

Well I do have time so I can configure one server with more than physical cores and the other limited to the physical cores and compare RAC.

[leadeater]

Right now on the DL560's I'm testing 1 VM with 32 vCPU vs 4 VMs with 8 vCPU each. To see if running each VM on their own NUMA node is overall faster in total. Maybe not as each task is a single CPU thread so I doubt there is any performance loss due to NUMA domain crossing (shouldn't happen at all). I assume this is the case but I'd really like to actually know 

[Windows7ge]

45 minutes ago, leadeater said:

Can't most of them are VMs and you actually get a bit of performance regression if you create VMs greater than the total actual number of cores.

 

Well I do have time so I can configure one server with more than physical cores and the other limited to the physical cores and compare RAC.

Interesting. I'd just run BOINC on the host kernel. I've not experimented with ESXi or vSphere as I think they're calling it now but from my testing running BOINC in a VM causes noticeable loss of compute. We want as little overhead as possible especially for our big players.

 

34 minutes ago, leadeater said:

Right now on the DL560's I'm testing 1 VM with 32 vCPU vs 4 VMs with 8 vCPU each. To see if running each VM on their own NUMA node is overall faster in total. Maybe not as each task is a single CPU thread so I doubt there is any performance loss due to NUMA domain crossing (shouldn't happen at all). I assume this is the case but I'd really like to actually know 

I concur. My expectation would be little to no difference. Worth doing a real world test on to validate though.

[Critaek]

Hi everyone, joined the team to help!

I need some help. I downloaded some activity, and i have 7 total (stopped downloading, i just wanted to test for now). I have 6 running but my cpu is only at 80%, with space to accomodate another one running, but boinc just runs 6. Is this normal? Also i have 100% usage set in the SiDock@home site so it should use all of the avaiable cpu power!

[GOTSpectrum]

Let's hope they actually do the shout out with the correct link 

 

[Lightwreather]

2 hours ago, GOTSpectrum said:

Let's hope they actually do the shout out with the correct link 

 

Oh so, you're spectrum

[GOTSpectrum]

Just now, J-from-Nucleon said:

Oh so, you're spectrum

Yeah that's me! The Folding team leader. @Ithanulis the BOINC team leader and usually handles most of this stuff but they have been really busy this year so I though I'd lend a hand. 

[leadeater]

5 hours ago, Windows7ge said:

I've not experimented with ESXi or vSphere as I think they're calling it now but from my testing running BOINC in a VM causes noticeable loss of compute.

Sort of yes and also no, it's possible to not have a big impact but there is a ton of tweaks and optimizations that you need to do. VMware has a bunch of HPC recommendations and whitepapers but BOINC is quite different from actual HPC. The systems running VMs just have to be VMs sadly as I'm using unused resource capacity so it is what it is.

 

That's why I'm hoping the 8 vCPU VMs running on a single NUMA domain each will result in less performance impact, only time will tell on that one.

 

The 5x 12 core systems are baremetal and I'll have others too soon so not all bad. Haven't even started the setup of the systems I'll be running at home yet (not that this will be nearly as much).

[Windows7ge]

3 minutes ago, leadeater said:

Sort of yes and also no, it's possible to not have a big impact but there is a ton of tweaks and optimizations that you need to do. VMware has a bunch of HPC recommendations and whitepapers but BOINC is quite different from actual HPC. The systems running VMs just have to be VMs sadly as I'm using unused resource capacity so it is what it is.

 

That's why I'm hoping the 8 vCPU VMs running on a single NUMA domain each will result in less performance impact, only time will tell on that one.

 

The 5x 12 core systems are baremetal and I'll have others too soon so not all bad. Haven't even started the setup of the systems I'll be running at home yet (not that this will be nearly as much).

I can sort of agree with this circumstantially. Working with QEMU/KVM there are some edits to make to the default config that give you that closer to bare metal performance. Part of the issue here though is if the host happens to do something while most of the cores are already pegged you can watch their output dip hard.

 

You pinned them right? Otherwise AFAIK if vSphere handles it like QEMU/KVM it will look as the physical cores like a pool and hand them out without regard for which physical socket they're on. I think this brings us back to NUMA vs UMA and weather that matters or not given the work isn't technically multi-threaded.

 

I would tell you to just spin-up a quick'n'dirty Linux install on USB and use that but running BOINC on Linux from USB is exactly what appears to be what was causing my issues on the EPYC 7601 32C/64T. Something about that USB wasn't allowing the CPU to do it's job and crunch WU's. We are at the 45.5 hour mark running on a super random WD 3TB HDD and we're still pegged at ~61.00/64.00.

[leadeater]

1 hour ago, Windows7ge said:

You pinned them right? Otherwise AFAIK if vSphere handles it like QEMU/KVM it will look as the physical cores like a pool and hand them out without regard for which physical socket they're on

Nah ESXi is properly NUMA aware and knows to keep threads on same NUMA node if communicating with each other. It can even detect different VMs communicating with each other and will preference to keep them on the same NUMA  node as well.

 

ESXi also has a latency sensitive flag you can put on a VM and if you do that the cores assigned to that VM are exclusive to it, it bypasses the scheduler. It also means you can no longer share those cores of course, and you need to leave one for ESXi itself.

[Windows7ge]

1 minute ago, leadeater said:

Nah ESXi is properly NUMA aware and knows to keep threads on same NUMA node if communicating with each other. It can even detect different VMs communicating with each other and will preference to keep them on the same NUMA  node as well.

Ah, nah, QEMU/KVM you have to tell it to do that. Doesn't do it on it's own.

 

Well post back with the results when you have them. It will be something worth noting for the future.

[leadeater]

2 minutes ago, Windows7ge said:

Ah, nah, QEMU/KVM you have to tell it to do that. Doesn't do it on it's own.

 

Well post back with the results when you have them. It will be something worth noting for the future.

Yea, I'm still sure it's slower than baremetal as I don't really know all the secrets to get the maximum performance for this situation, but it's this or not being able to use the hardware at all and something is better than nothing lol