[Beowulf] ***UNCHECKED*** Re: [EXTERNAL] Re: Re: Spark, Julia, OpenMPI etc. - all in one place

[Lux, Jim (US 7140)]

What I find fascinating about the poster from NASA is the comment about "man in the loop data product generation" - this is what has always interested me - being able to get interactive supercomputing - My desires have always been to run moderately large models (run time >30 seconds) with large parameter spaces, in an interactive sense, so I can "turn the knobs" on the design. 

As a result, my particular interests have run more towards the embarrassingly parallel, run lots of cases that a single node can handle, with the associated scatter/gather.  This is opposed to running "big models" (although I've had reasons to do that).

For example, in the last year, I've been running a lot of models of an antenna forming the Owens Valley Radio Observatory Long Wavelength Array (OVRO-LWA). This is a large array of hundreds of antennas scattered across a few sq km near Big Pine, CA and observes the cosmos in the 30-80 MHz band. The properties of a single antenna are easy and quick to model.  But there's a bunch of questions that require more time - What's the interaction between the antennas? How close can they be and not interact? What's the effect of manufacturing tolerances? When it rains, and the dirt under the antenna is wet, how much does that change the response?

Similarly, I've been doing models of wire antennas on the surface of the Moon, for 100kHz to 50 MHz.  Any one antenna is trivial and quick to model (and, for that matter, there are analytic models that are pretty good).  But we've got the same questions.  What if the rover laying the wire out doesn't do it in a perfectly straight line? What's the interaction between 2 antennas that are 300 meters apart (given that the wavelength at 100kHz is 3km, the antennas are "close" in electromagnetic terms)? 

These are really sort of Monte Carlo type analyses (much like running multiple runs of weather models with slightly different starting parameters).

Some time in the past (>10 years ago), I was really interested in "field supercomputing" - there are problems (subsurface imaging by ground penetrating radar) that require a lot of computation.  And you don't have a fat pipe to a HPC facility to send the data (satphones at 10s of kbps is the notional scenario).  But here, you want something that will survive field use - no computer room, preferably a sealed box, etc. 

Interestingly, all of these are "personal HPC" - that is, the goal is to have a HPC capability that is controlled locally - you don't have to compete in a queue for resources, etc. and because it's "local", there's no shoveling data to the HPC and getting results back.  Further, it's interactive - you want to tweak, run it again, and get the answers back quickly - single digit minutes at most.  This has been described as high throughput computing, but I'm not sure that's right - to me that implies sustained bandwidth - it's kind of like "station wagon full of tapes" has high data throughput but long latency.  Latency is important for human scale interaction.

These days, data pipes are easier to come by - A field scientist or engineer could send megabits to a remote HPC center.  I send my data and computation to TACC, in Texas, and half of JPL's latest cluster Gattaca is hosted at SuperNAP, hundreds of miles from my office.  But neither of those are truly interactive - there are batch queues, and while my jobs are small enough to get done quickly (minutes), it *is* redolent of when I was a youth, submitting my deck over the counter and coming back a few hours later to pick up the greenbar paper.  I want that immediacy - glowing digits and characters on the screen appearing instantaneously, even if it's "divide by zero error in Line 310", as opposed to sitting at the keypunch and standing in line.  (I do have to chuckle, though, at my CS lecturers in the late 1970s making a big deal about desk checking your code before submitting, because time is money, and computer time is more expensive than your time - how life has changed in 40 years).





On 10/15/20, 7:15 AM, "Beowulf on behalf of Michael Di Domenico" <beowulf-bounces at beowulf.org on behalf of mdidomenico4 at gmail.com> wrote:

    ah, interesting.

    this is what i was referring to, which is what i believe is codified
    in the "beowulf" book i recall reading.

    https://urldefense.us/v3/__https://spinoff.nasa.gov/Spinoff2020/it_1.html__;!!PvBDto6Hs4WbVuu7!bIrJokEF-PV-9u42csgECfdGQv6CPxfH654QEeVT__BsDh4PURoz-fJxq23GkgQ_OaRx87I$ 

    but it seems this also exists.  i can't recall offhand whether that
    was mentioned in the book or not

    https://urldefense.us/v3/__https://www.hq.nasa.gov/hpcc/reports/annrpt97/accomps/ess/WW49.html__;!!PvBDto6Hs4WbVuu7!bIrJokEF-PV-9u42csgECfdGQv6CPxfH654QEeVT__BsDh4PURoz-fJxq23GkgQ_c8ZFY-A$ 


    On Thu, Oct 15, 2020 at 9:44 AM Douglas Eadline <deadline at eadline.org> wrote:
    >
    >
    > > On Thu, Oct 15, 2020 at 12:10 AM Lux, Jim (US 7140) via Beowulf
    > > <beowulf at beowulf.org> wrote:
    > >>
    > >> Well, maybe a Beowulf cluster of yugos…
    > >
    > > not really that far of a stretch, from what i can recall wasn't the
    > > first beowulf cluster a smattering of random desktops layout on the
    > > floor in an office
    >
    > Actually it was a single small cabinet with 486 processor
    > motherboards and 10Mbit Ethernet with a hub. There is
    > a small picture of it on the SC14 Beowulf Bash invite
    > (in the middle) As I recall we could only find an old
    > small picture of it.
    >
    > https://urldefense.us/v3/__https://www.clustermonkey.net/Supercomputing/beowulf-bash-invitations-2008-to-present.html__;!!PvBDto6Hs4WbVuu7!bIrJokEF-PV-9u42csgECfdGQv6CPxfH654QEeVT__BsDh4PURoz-fJxq23GkgQ_zWEo5Pw$ 
    >
    > From there all kinds of configurations appeared.
    > Including mostly "workstations" on wire shelves
    > (the differentiation between "desktop" and "server"
    > was just starting with the introduction of the Pentium-Pro)
    >
    > For those interested in the Beowulf history you can watch
    > the short video (fully shareable BTW, sponsored by AMD)
    >
    >   https://urldefense.us/v3/__https://www.youtube.com/watch?v=P-epcSlAFvI__;!!PvBDto6Hs4WbVuu7!bIrJokEF-PV-9u42csgECfdGQv6CPxfH654QEeVT__BsDh4PURoz-fJxq23GkgQ_FyrFvOo$ 
    >
    >
    > --
    > Doug
    >
    > > _______________________________________________
    > > Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
    > > To change your subscription (digest mode or unsubscribe) visit
    > > https://urldefense.us/v3/__https://beowulf.org/cgi-bin/mailman/listinfo/beowulf__;!!PvBDto6Hs4WbVuu7!bIrJokEF-PV-9u42csgECfdGQv6CPxfH654QEeVT__BsDh4PURoz-fJxq23GkgQ_or4lTPE$ 
    > >
    >
    >
    > --
    > Doug
    >
    _______________________________________________
    Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
    To change your subscription (digest mode or unsubscribe) visit https://urldefense.us/v3/__https://beowulf.org/cgi-bin/mailman/listinfo/beowulf__;!!PvBDto6Hs4WbVuu7!bIrJokEF-PV-9u42csgECfdGQv6CPxfH654QEeVT__BsDh4PURoz-fJxq23GkgQ_or4lTPE$