<div dir="ltr"><p class="MsoNormal" style="margin-bottom:12pt"><span>> All of a sudden
simple “send the same command to all nodes” just doesn’t work. And
that’s what will inevitably be the case as we scale up in the HPC world –
there
will always be dead or malfunctioning nodes.<u></u><u></u></span></p>
<p class="MsoNormal" style="margin-bottom:12pt"><span>Jim, this is true. And 'we' should be looking to the webscale generation for the answers. They thought about computing at scale from the beginning.</span></p><p class="MsoNormal" style="margin-bottom:12pt"><span>Regarding hardware failures, I heard a shaggy dog story that Microsoft/Amazon/Google order servers ready racked in shipping containers. When a certain proportion of servers are dead, they simply close it down and move on.<br>Can anyone confirm or deny this story? <br></span></p><p class="MsoNormal" style="margin-bottom:12pt"><span> Which brings me to another one of my hobby horses - the environmental costs of HPC. When pitching HPC clusters you often put in an option for a mid-life upgrade. I think upping the RAM is quite common, but processors and interconnect much less so.<br></span></p><p class="MsoNormal" style="margin-bottom:12pt"><span>So kit is hopefulyl worked hard for five years, till the cost of power and cooling is outweighed by the performance of a new generation. But where does the kit get recycled? Again when pitching clusters you have to put in guarantees about WEE (or the equivalent in the USA)<br></span></p><br></div><div class="gmail_extra"><br><div class="gmail_quote">On 8 May 2018 at 18:34, Lux, Jim (337K) <span dir="ltr"><<a href="mailto:james.p.lux@jpl.nasa.gov" target="_blank">james.p.lux@jpl.nasa.gov</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div link="blue" vlink="purple" lang="EN-US">
<div class="m_-3115792153076278501WordSection1">
<p class="MsoNormal"><u></u> <u></u></p>
<p class="MsoNormal"><u></u> <u></u></p>
<div style="border:none;border-top:solid #b5c4df 1.0pt;padding:3.0pt 0in 0in 0in">
<p class="MsoNormal"><b><span style="font-size:12.0pt;color:black">From: </span></b><span style="font-size:12.0pt;color:black">Beowulf <<a href="mailto:beowulf-bounces@beowulf.org" target="_blank">beowulf-bounces@beowulf.org</a>> on behalf of "<a href="mailto:beowulf@beowulf.org" target="_blank">beowulf@beowulf.org</a>" <<a href="mailto:beowulf@beowulf.org" target="_blank">beowulf@beowulf.org</a>><br>
<b>Reply-To: </b>John Hearns <<a href="mailto:hearnsj@googlemail.com" target="_blank">hearnsj@googlemail.com</a>><br>
<b>Date: </b>Thursday, May 3, 2018 at 6:54 AM<br>
<b>To: </b>"<a href="mailto:beowulf@beowulf.org" target="_blank">beowulf@beowulf.org</a>" <<a href="mailto:beowulf@beowulf.org" target="_blank">beowulf@beowulf.org</a>><br>
<b>Subject: </b>Re: [Beowulf] Bright Cluster Manager<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><u></u> <u></u></p>
</div>
<div><div><div class="h5">
<div>
<div>
<div>
<div>
<div>
<p class="MsoNormal"><a name="m_-3115792153076278501__MailOriginalBody">I agree with Doug. The way forward is a lightweight OS with containers for the applications.<u></u><u></u></a></p>
</div>
<p class="MsoNormal"><span>I think we need to learn from the new kids on the block - the webscale generation.<u></u><u></u></span></p>
</div>
<p class="MsoNormal"><span>They did not go out and look at how massive supercomputer clusters are put together.<u></u><u></u></span></p>
</div>
<p class="MsoNormal"><span>No, they went out and build scale out applications built on public clouds.<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>We see 'applications designed to fail' and 'serverless'<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span><u></u> <u></u></span></p>
</div>
<p class="MsoNormal"><span>Yes, I KNOW that scale out applications like these are Web type applications, and all application examples you
<u></u><u></u></span></p>
</div>
<p class="MsoNormal"><span>see are based on the load balancer/web server/database (or whatever style) paradigm<u></u><u></u></span></p>
</div></div><div><div><div class="h5">
<div>
<p class="MsoNormal"><span><u></u> <u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>The art of this will be deploying the more tightly coupled applications with HPC has,<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span>which depend upon MPI communications over a reliable fabric, which depend upon GPUs etc.<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>The other hat I will toss into the ring is separating parallel tasks which require computation on several<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>servers and MPI communication between them versus 'embarrassingly parallel' operations which may run on many, many cores<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span>but do not particularly need communication between them.<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>The best successes I have seen on clusters is where the heavy parallel applications get exclusive compute nodes.<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>Cleaner, you get all the memory and storage bandwidth and easy to clean up. Hell, reboot the things after each job. You got an exclusive node.<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>I think many designs of HPC clusters still try to cater for all workloads - Oh Yes, we can run an MPI weather forecasting/ocean simulation<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span>But at the same time we have this really fast IO system and we can run your Hadoop jobs.
<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>I wonder if we are going to see a fork in HPC. With the massively parallel applications being deployed, as Doug says, on specialised
<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>lightweight OSes which have dedicated high speed, reliable fabrics and with containers.<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>You won't really be able to manage those systems like individual Linux servers. Will you be able to ssh in for instance?<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>ssh assumes there is an ssh daemon running. Does a lightweight OS have ssh? Authentication Services? The kitchen sink?<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span><u></u> <u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>The less parallel applications being run more and more on cloud type installations, either on-premise clouds or public clouds.<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>I confound myself here, as I cant say what the actual difference between those two types of machines is, as you always needs<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>an interconnect fabric and storage, so why not have the same for both types of tasks.<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>Maybe one further quip to stimulate some conversation. Silicon is cheap. No, really it is.<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal"><span>Your friendly Intel salesman may wince when you say that. After all those lovely Xeon CPUs cost north of 1000 dollars each.<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span>But again I throw in some talking points:<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span>power and cooling costs the same if not more than your purchase cost over several years<u></u><u></u></span></p>
</div>
</div></div><div><div><div class="h5">
<div style="border:none;border-bottom:solid windowtext 1.0pt;padding:0in 0in 1.0pt 0in">
<p class="MsoNormal" style="margin-bottom:12.0pt;border:none;padding:0in"><span>are we exploiting all the capabilities of those Xeon CPUs<br>
<br>
<u></u><u></u></span></p>
</div>
</div></div><p class="MsoNormal" style="margin-bottom:12.0pt"><span>And another aspect of this - I’ve been doing stuff with “loose clusters” of low capability processors (Arduino, Rpi, Beagle) doing distributed sensing kinds of tasks
– leaving aside the Arduino (no OS) – the other two wind up with some flavor of Debian but often with lots of stuff you don’t need (i.e. Apache). Once you’ve fiddled with one node to get the configuration right, you want to replicate it across a bunch of nodes
– right now that means sneakernet of SD cards - although in theory, one should be able to push an image out to the local file system (typically 4GB eMMC in the case of beagles), and tell it to write that to the “boot area” – but I’ve not tried it.
<u></u><u></u></span></p>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span>While I’d never claim my pack of beagles is HPC, it does share some aspects – there’s parallel work going on, the nodes need to be aware of each other and synchronize
their behavior (that is, it’s not an embarrassingly parallel task that’s farmed out from a queue), and most importantly, the management has to be scalable. While I might have 4 beagles on the bench right now – the idea is to scale the approach to hundreds.
Typing “sudo apt-get install tbd-package” on 4 nodes sequentially might be ok (although pdsh and csshx help a lot) it’s not viable for 100 nodes.<u></u><u></u></span></p>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span>The other aspect of my application that’s interesting, and applicable to exascale kinds of problems, is tolerance to failures – if I have a low data rate link among
nodes (with not necessarily all to all connectivity), one can certainly distribute a new OS image (or container) with time. There’s some ways to deal with errors in the transfers (other than just retransmit all – which doesn’t work if the error rate is high
enough that you can guarantee at least one error will occur in a long transfer). But how do you *<b>manage</b>* a cluster with hundreds or thousands of nodes where some fail randomly, reset randomly, etc.<u></u><u></u></span></p>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span>All of a sudden simple “send the same command to all nodes” just doesn’t work. And that’s what will inevitably be the case as we scale up in the HPC world – there
will always be dead or malfunctioning nodes.<u></u><u></u></span></p>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span><u></u> <u></u></span></p>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span><u></u> <u></u></span></p>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span><u></u> <u></u></span></p>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span><u></u> <u></u></span></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span><br>
<br>
<br>
<br>
<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span><br>
<br>
<br>
<br>
<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span><br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span><br>
<br>
<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span><br>
<br>
<br>
<br>
<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span><br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span><br>
<br>
<br>
<br>
<br>
<u></u><u></u></span></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span><u></u> <u></u></span></p>
<div>
<div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><span><br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<u></u><u></u></span></p>
</div>
</div>
</div>
</div>
</div>
</div><div><div class="h5">
<div>
<p class="MsoNormal"><span><u></u> <u></u></span></p>
<div>
<p class="MsoNormal"><span>On 3 May 2018 at 15:04, Douglas Eadline <</span><a href="mailto:deadline@eadline.org" target="_blank"><span>deadline@eadline.org</span><span></span></a><span>>
wrote:<u></u><u></u></span></p>
<blockquote style="border:none;border-left:solid #cccccc 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<p class="MsoNormal"><span><br>
<br>
Here is where I see it going<br>
<br>
1. Computer nodes with a base minimal generic Linux OS<br>
(with PR_SET_NO_NEW_PRIVS in kernel, added in 3.5)<br>
<br>
2. A Scheduler (that supports containers)<br>
<br>
3. Containers (Singularity mostly)<br>
<br>
All "provisioning" is moved to the container. There will be edge cases of<br>
course, but applications will be pulled down from<br>
a container repos and "just run"<br>
<br>
--<br>
Doug<br>
<br>
<br>
> I never used Bright. Touched it and talked to a salesperson at a<br>
> conference but I wasn't impressed.<br>
><br>
> Unpopular opinion: I don't see a point in using "cluster managers"<br>
> unless you have a very tiny cluster and zero Linux experience. These<br>
> are just Linux boxes with a couple applications (e.g. Slurm) running on<br>
> them. Nothing special. xcat/Warewulf/Scyld/Rocks just get in the way<br>
> more than they help IMO. They are mostly crappy wrappers around free<br>
> software (e.g. ISC's dhcpd) anyway. When they aren't it's proprietary<br>
> trash.<br>
><br>
> I install CentOS nodes and use<br>
> Salt/Chef/Puppet/Ansible/<wbr>WhoCares/Whatever to plop down my configs and<br>
> software. This also means I'm not suck with "node images" and can<br>
> instead build everything as plain old text files (read: write SaltStack<br>
> states), update them at will, and push changes any time. My "base<br>
> image" is CentOS and I need no "baby's first cluster" HPC software to<br>
> install/PXEboot it. YMMV<br>
><br>
><br>
> Jeff White<br>
><br>
> On 05/01/2018 01:57 PM, Robert Taylor wrote:<br>
>> Hi Beowulfers.<br>
>> Does anyone have any experience with Bright Cluster Manager?<br>
>> My boss has been looking into it, so I wanted to tap into the<br>
>> collective HPC consciousness and see<br>
>> what people think about it.<br>
>> It appears to do node management, monitoring, and provisioning, so we<br>
>> would still need a job scheduler like lsf, slurm,etc, as well. Is that<br>
>> correct?<br>
>><br>
>> If you have experience with Bright, let me know. Feel free to contact<br>
>> me off list or on.<br>
>><br>
>><br>
>><br>
>> ______________________________<wbr>_________________<br>
>> Beowulf mailing list, </span><a href="mailto:Beowulf@beowulf.org" target="_blank"><span>Beowulf@beowulf.org</span><span></span></a><span> sponsored by Penguin
Computing<br>
>> To change your subscription (digest mode or unsubscribe) visit<br>
>> </span><a href="https://urldefense.proofpoint.com/v2/url?u=http-3A__www.beowulf.org_mailman_listinfo_beowulf&d=DwIGaQ&c=C3yme8gMkxg_ihJNXS06ZyWk4EJm8LdrrvxQb-Je7sw&r=DhM5WMgdrH-xWhI5BzkRTzoTvz8C-BRZ05t9kW9SXZk&m=2km_EqLvNf2v9rNf8LphAYkJ-Sc_azfEyHqyDIzpLOc&s=kq0wdhy80VqcBCwcQAAQa0RbsgWIekhd0qU0zC81g1Q&e=" target="_blank"><span>https://urldefense.proofpoint.<wbr>com/v2/url?u=http-3A__www.<wbr>beowulf.org_mailman_listinfo_<wbr>beowulf&d=DwIGaQ&c=<wbr>C3yme8gMkxg_<wbr>ihJNXS06ZyWk4EJm8LdrrvxQb-<wbr>Je7sw&r=DhM5WMgdrH-<wbr>xWhI5BzkRTzoTvz8C-<wbr>BRZ05t9kW9SXZk&m=2km_<wbr>EqLvNf2v9rNf8LphAYkJ-Sc_<wbr>azfEyHqyDIzpLOc&s=<wbr>kq0wdhy80VqcBCwcQAAQa0RbsgWIek<wbr>hd0qU0zC81g1Q&e=</span><span></span></a><span><br>
><br>
><br>
> --<br>
> MailScanner: Clean<br>
><br>
> ______________________________<wbr>_________________<br>
> Beowulf mailing list, </span><a href="mailto:Beowulf@beowulf.org" target="_blank"><span>Beowulf@beowulf.org</span><span></span></a><span> sponsored by Penguin
Computing<br>
> To change your subscription (digest mode or unsubscribe) visit<br>
> </span><a href="http://www.beowulf.org/mailman/listinfo/beowulf" target="_blank"><span>http://www.beowulf.org/<wbr>mailman/listinfo/beowulf</span><span></span></a><span><br>
><br>
<br>
<br>
-- <br>
Doug<br>
<br>
-- <br>
MailScanner: Clean<br>
<br>
______________________________<wbr>_________________<br>
Beowulf mailing list, </span><a href="mailto:Beowulf@beowulf.org" target="_blank"><span>Beowulf@beowulf.org</span><span></span></a><span> sponsored by Penguin
Computing<br>
To change your subscription (digest mode or unsubscribe) visit </span><a href="http://www.beowulf.org/mailman/listinfo/beowulf" target="_blank"><span>http://www.beowulf.org/<wbr>mailman/listinfo/beowulf</span><span></span></a><span><u></u><u></u></span></p>
</blockquote>
</div>
<p class="MsoNormal"><span><u></u> <u></u></span></p>
</div>
</div></div></div>
</div>
<br>______________________________<wbr>_________________<br>
Beowulf mailing list, <a href="mailto:Beowulf@beowulf.org">Beowulf@beowulf.org</a> sponsored by Penguin Computing<br>
To change your subscription (digest mode or unsubscribe) visit <a href="http://www.beowulf.org/mailman/listinfo/beowulf" rel="noreferrer" target="_blank">http://www.beowulf.org/<wbr>mailman/listinfo/beowulf</a><br>
<br></blockquote></div><br></div>