[Thinlinc-technical] [External] Using fair share scheduling on ThinLinc agent hosts

[Thota, Abhinav]

Hello Bjorn, All,

We are doing something similar, also with Cgroups, you can check it out here: https://github.com/rperigo/cgroup_py <https://github.com/rperigo/cgroup_py>

Best,
Abhinav 

--
Abhinav Thota
athota at iu.edu | 812-855-3701
Team Lead, Scientific Applications and Performance Tuning (SciAPT)
Research Technologies
Indiana University

> On Oct 10, 2018, at 8:10 AM, Björn Fischer <bf at CeBiTec.Uni-Bielefeld.DE> wrote:
> 
> This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources.
> -------
> 
> Dear subscribers,
> 
> at our site we are currently experimenting with fair share scheduling on
> ThinLinc agent hosts. Our approach still has some issues, but generally
> it looks quite promising so that we would like to share our experience
> with all of you. Any feedback on this topic is highly appreciated.
> 
> At the Center for Biotechnology (CeBiTec) at Bielefeld University we
> provide scientist's desktops (internet and office applications, and
> scientific applications) for researchers and teachers.
> 
> Formerly based on SunRay technology (using Solaris' FSS), we now use a
> ThinLinc-4.9 setup with 5 VSM agents (each 56 Cores, 376GB RAM). Our
> approach uses Ubuntu-18.04 on the bare metal and each VSM agent is
> running inside an LXD-3.0 container, also with an Ubuntu-18.04 user
> space. This setup works exceedingly well wrt migration and deployment
> of agent hosts.
> 
> Usually, about 30 users share a single VSM agent host which handles
> the load quite well -- as long as there is no firefox, thunderbird, or
> similar process going berserk. It may also occur that an interactive
> science app suddenly generates lots of load. These situations lead
> to lags for all users on the system deteriorating the general user
> experience.
> 
> In order to fix that problem we are experimenting with the fair share
> scheduling feature of the Linux kernel which is controlled by the
> cgroups API. With fair share scheduling enabled, a user can acquire all
> CPU resources he/she wants as long as there is idle time on the system.
> When CPU time is exhausted the scheduler splits CPU time equally across
> all demanding users. In other words the user cannot single-handedly slow
> down the whole system any more.
> 
> The deployment we have at hand right now seems to work quite well apart
> from some issues with systemd.
> 
> Is anyone doing something similar?
> 
> If anyone is interested in technical details of our setup, or may
> provide help with our systemd problems, please let me know.
> 
> Björn
> -- 
> CeBiTec -- Bioinformatics Resource Facility
> _______________________________________________
> Thinlinc-technical mailing list
> Thinlinc-technical at lists.cendio.se
> Manage your subscription:
> http://lists.cendio.se/mailman/listinfo/thinlinc-technical

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.cendio.se/pipermail/thinlinc-technical/attachments/20181010/86578bc8/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 4031 bytes
Desc: not available
URL: <http://lists.cendio.se/pipermail/thinlinc-technical/attachments/20181010/86578bc8/attachment.p7s>