[Beowulf] MPI, fault handling, etc.
Lux, Jim (337C)
james.p.lux at jpl.nasa.gov
Fri Mar 11 13:28:16 PST 2016
I’ll agree that architecture is the key.
Packet switching breaks when you need to move petabytes in a reasonable time across large distances, because each intermediate node has to have storage for DataRate*RoundTripTime (so that flow control works). If you’re moving Tbits/sec (e.g. something like 3d medical images) across the country, that means your store and forward switches need to store a fair amount of data.
That’s a narrow example..
But I’m not sure you can always trade things to get more reliability without impacting performance.
At some point, you can probably take an information theory approach. You’ve got a channel with a certain error rate, and you want to push a certain data rate through it at a particular (lower) error rate.
From: Justin Y. Shi [mailto:shi at temple.edu]
Sent: Thursday, March 10, 2016 3:12 PM
To: Lux, Jim (337C) <james.p.lux at jpl.nasa.gov>
Cc: Douglas Eadline <deadline at eadline.org>; beowulf at beowulf.org
Subject: Re: [Beowulf] MPI, fault handling, etc.
Not only high reliability is possible, incrementally higher performance is also possible at the same time or the Internet would crumble by now.
The key is in the architecture. The overheads can be traded for performance gains elsewhere without sacrificing reliability or scalability.
In other words, the packet switching network proved that you can have the cake and eat it at the same time.
On Thu, Mar 10, 2016 at 4:43 PM, Lux, Jim (337C) <james.p.lux at jpl.nasa.gov<mailto:james.p.lux at jpl.nasa.gov>> wrote:
High reliability is possible, at the expense of substantial additional resources. Packet switching requires storage to hold unacknowledged packets that might need to be resent and it requires time for the resends.
If you have time critical processes with fine granularity, then you have to budget in the memory and time to achieve that reliability.
The other thing is that some high reliability approaches do not scale well. A bit error rate of 1E-9 would be outstandingly good on a radio link, but if you’re sending data at 1Gbps, that’s an error every second. If you’re sending data at 1 Tbps, that’s an error every millisecond. If your packets are a millisecond long, then you get no successful packets through.
Fortunately, “over the wire” error rates are much better (1E-12 wouldn’t be unusual), but that’s still a once every 20 minutes sort of error rate at 1 Gbps. And if you’re talking about interprocessor interconnects.. a 12x Infiniband is about 300 Gb/sec. A 1E-12 error rate would be an error every few seconds. At those rates, of course you’re going to use some sort of forward error correction (you’re not going to depend on retry), and that might be an effective data rate reduction of 30-50% (depending on the coding).
From: Justin Y. Shi [mailto:shi at temple.edu<mailto:shi at temple.edu>]
Sent: Thursday, March 10, 2016 1:14 PM
To: Douglas Eadline <deadline at eadline.org<mailto:deadline at eadline.org>>
Cc: Lux, Jim (337C) <james.p.lux at jpl.nasa.gov<mailto:james.p.lux at jpl.nasa.gov>>; beowulf at beowulf.org<mailto:beowulf at beowulf.org>
Subject: Re: [Beowulf] MPI, fault handling, etc.
Not that fast though. 100% reliability is practically achievable. And we are enjoying the results everyday. I mean the wireless and wired packet switching networks.
The problem is our tendency of drawing fast conclusions. The one human made architecture that defies the "curse" of component failure is the statistic multiplexing principle (or packet switching).
It has proven to work perfectly using growing number of not-so-reliable devises without suffering the scalability dilemma. We should learn how to apply that technology to extreme scale computing. To this day, the full extends of protocol logics still cannot be adequately described formally on paper. But they work well, if done right.
On Thu, Mar 10, 2016 at 3:44 PM, Douglas Eadline <deadline at eadline.org<mailto:deadline at eadline.org>> wrote:
> I will supper C's "hater" listing effort just to keep a spot light on the
> important subject.
> The question is not MPI is efficient or not. Fundamentally, all
> will fail in unexpected ways. Bare metal computing was important decades
> ago but detrimental to large scale computing. It is simply flawed for
> extreme scale computing.
> The Alan Fekete, Nancy Lynch, John Spinneli's impossible proof is the
> fundamental "line in the sand" that cannot be crossed.
> The corollary of that proof is that it is impossible to detect failure
> reliably either. Therefore, those efforts for for runtime
> detection/repair/reschedule are also flawed for extreme scale computing.
Well on that note, I suppose we should just call it day.
Although, some thought Godel would put the whole math thing
out of business as well.
> On Thu, Mar 10, 2016 at 8:44 AM, Lux, Jim (337C)
> <james.p.lux at jpl.nasa.gov<mailto:james.p.lux at jpl.nasa.gov>>
>> This is interesting stuff.
>> Think back a few years when we were talking about checkpoint/restart
>> issues: as the scale of your problem gets bigger, the time to checkpoint
>> becomes bigger than the time actually doing useful work.
>> And, of course, the reason we do checkpoint/restart is because it’s
>> bare-metal and easy. Just like simple message passing is “close to
>> metal” and “straightforward”.
>> Similarly, there’s “fine grained” error detection and correction:
>> codes in memory; redundant comm links or retries. Each of them imposes
>> some speed/performance penalty (it takes some non-zero time to compute
>> syndrome bits in a ECC, and some non-zero time to fix the errored
>> bits… in
>> a lot of systems these days, that might be buried in a pipeline, but the
>> delay is there, and affects performance)
>> I think of ECC as a sort of diffuse fault management: it’s pervasive,
>> uniform, and the performance penalty is applied evenly through the
>> Redundant (in the TMR sense) links are the same way.
>> Retries are a bit different. The “detecting” a fault is diffuse and
>> pervasive (e.g. CRC checks occur on each message), but the correction of
>> the fault is discrete and consumes resources at that time. In a system
>> with tight time coupling (a pipelined systolic array would be the sort
>> worst case), many nodes have to wait to fix the one that failed.
>> A lot depends on the application: tighter time coupling is worse than
>> embarrassingly parallel (which is what a lot of the “big data” stuff
>> fundamentally EP, scatter the requests, run in parallel, gather the
>> The challenge is doing stuff in between: You may have a flock with
>> capacity (just as ECC memory might have 1.5N physical storage bits to be
>> used to store N bits), but how do you automatically distribute the
>> resources to be failure tolerant. The original post in the thread
>> out that MPI is not a particularly facile tool for doing this. But
>> I’m not
>> sure that there is a tool, and I’m not sure that MPI is the root of
>> lack of tools. I think it’s that moving from close to the metal is
>> “hard problem” to do in a generic way. (The issues about 32 bit
>> counts are
>> valid, though)
>> James Lux, P.E.
>> Task Manager, DHFR Space Testbed
>> Jet Propulsion Laboratory
>> 4800 Oak Grove Drive, MS 161-213
>> Pasadena CA 91109
>> +1(818)395-2714<tel:%2B1%28818%29395-2714> (cell)
>> Beowulf mailing list, Beowulf at beowulf.org<mailto:Beowulf at beowulf.org> sponsored by Penguin Computing
>> To change your subscription (digest mode or unsubscribe) visit
> Mailscanner: Clean
> Beowulf mailing list, Beowulf at beowulf.org<mailto:Beowulf at beowulf.org> sponsored by Penguin Computing
> To change your subscription (digest mode or unsubscribe) visit
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Beowulf