river-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Gregg Wonderly <ge...@cox.net>
Subject Re: tree-based / log(p) communication cost algorithms for River?
Date Sun, 13 Sep 2015 04:18:27 GMT
I guess without more knowledge of what your problem space really is I can’t understand why
the problem doesn’t decompose into small problems which can be farmed out to a java space.
 Certainly, there are issues if you are using a C-language application on and MDI platform
for parallelism.   If the barrier, right now, is the speed of the network vs the availability
of cycles on the GPU, then what you are doing now, probably makes since. But, if it would
be faster to infiniband a dozen requests out to other idle GPUs, then it would seem that there
might be something to be had by having a javaspace client grabbing requests and JNI calling
out to use the GPU for calculations.

Perhaps it feels comfortable now to use more GPU cycles close by and less infiniband traffic
for transport to other GPUs.  Whats the fraction describing the GPU time to compute a result
verse the infiniband time to transmit such data to another GPU?

What I am specifically wondering is what happens when one of the GPU machines crashes?  How
do you recover and continue processing?

Gregg Wonderly
 
> On Sep 11, 2015, at 4:22 AM, Bryan Thompson <bryan@systap.com> wrote:
> 
> Gregg,
> 
> Graphs traversal is in general a non-local problem with irregular, data
> dependent parallelism (the available fine grained parallelism for a vertex
> depends on its edge list, the size of the edge list can vary over many
> orders of magnitude, edges may connect vertices that are non local, and the
> size of the active frontier of vertices during traversal and also vary by
> many orders of magnitude).
> 
> The 2D decomposition minimizes the number of communications that need to be
> performed.  There are also hybrid decompositions that can minimize the
> communication volume (amount of data that needs to be transmitted),
> especially when combined with graph aware partitioning.
> 
> A data layout that ignores these issues will require more communication
> operations and/or have a greater communications volume.  Communication is
> the main barrier to scaling for graphs. So data layouts matter and
> innovation in data layouts is one of the key things driving improved
> scaling.
> 
> We actually use hardware acceleration as well. So each compute node has one
> or more gpus that are used to parallelize operations on the local edges.
> The gpus use an MPI distribution that uses RDMA over infiniband for off
> node communication. This avoids the overhead of having the CPU coordinate
> communication (data is not copied to the CPU but does directly over PCIe to
> the infiniband card).
> 
> The question about tree based communication patterns arises from an
> interest in being able to use java to coordinate data management
> activities.  These activities are not are as performance critical as the
> basic graph traversal, but they should still show good scaling. Tree based
> communication patterns are one way to achieve that scaling.
> 
> Thanks,
> Bryan
> 
> On Thursday, September 10, 2015, Gregg Wonderly <greggwon@gmail.com> wrote:
> 
>> Why doesn’t java spaces let you submit requests and have them worked on
>> and results returned without assigning any particular node any specific
>> responsibility?
>> 
>> Gregg
>> 
>>> On Aug 1, 2015, at 12:06 PM, Bryan Thompson <bryan@systap.com
>> <javascript:;>> wrote:
>>> 
>>> First, thanks for the responses and the interest in this topic.  I have
>>> been traveling for the last few days and have not had a chance to follow
>> up.
>>> 
>>> - The network would have multiple switches, but not multiple routers.
>> The
>>> typical target is an infiniband network.  Of course, Java let's us bind
>> to
>>> infiniband now.
>>> 
>>> - As far as I understand it, MPI relies on each node executing the same
>>> logic in a distributed communication pattern.  Thus the concept of a
>> leader
>>> election to determine a balanced tree probably does not show up. Instead,
>>> the tree is expressed in terms of the MPI rank assigned to each node.  I
>> am
>>> not suggesting that the same design pattern is used for river.
>>> 
>>> - We do need a means to define the relationship between a distributed
>>> communication pattern and the manner in which data are decomposed onto a
>>> cluster.  I am not sure that the proposal above gives us this directly,
>> but
>>> some extension of it probably would.   Let me give an example.  In our
>>> application, we are distributing the edges of a graph among a 2-D cluster
>>> of compute nodes (p x p machines).  The distribution is done by assigning
>>> the edges to compute nodes based on some function (key-range,
>>> hash-function) of the source and target vertex identifiers.  When we want
>>> to read all edges in the graph, we need to do an operation that is data
>>> parallel across either the rows (in-edges) or the columns (out-edges) of
>>> the cluster. See http://mapgraph.io/papers/UUSCI-2014-002.pdf for a TR
>> that
>>> describes this communication pattern for a p x p cluster of GPUs.  In
>> order
>>> to make this work with river, we would somehow have to associate the
>> nodes
>>> with their positions in this 2-D topology.  For example, we could
>> annotate
>>> each node with a "row" and "column" attribute that specifies its location
>>> in the compute grid.  We could then have a communicator for each row and
>>> each column based on the approach you suggest above.
>>> 
>>> The advantage of such tree based communication patterns is quite large.
>>> They require log(p) communication operations where you would otherwise
>> do p
>>> communication operations.  So, for example, only 4 communication
>> operations
>>> vs 16 for a 16 node cluster.
>>> 
>>> Thanks,
>>> Bryan
>>> 
>>> 
>>> On Wed, Jul 29, 2015 at 1:17 PM, Greg Trasuk <trasukg@stratuscom.com
>> <javascript:;>> wrote:
>>> 
>>>> 
>>>> I’ve wondered about doing this in the past, but for the workloads I’ve
>>>> worked with, I/O time has been relatively low compared to processing
>> time.
>>>> I’d guess there’s some combination of message frequency, cluster size
>> and
>>>> message size that makes it compelling.
>>>> 
>>>> The idea is interesting, though, because it could enable things like
>>>> distributed JavaSpaces, where we’d be distributing the search queries,
>> etc.
>>>> 
>>>> I would guess the mechanism would look like:
>>>> 
>>>> -Member nodes want to form a multicast group.
>>>> -They elect a leader
>>>> -Leader figures out a balanced notification tree, and passes it on to
>> each
>>>> member
>>>> -Leader receives multicast message and starts the message passing into
>> the
>>>> tree
>>>> -Recipients pass the message to local recipients, and also to their
>>>> designated repeater recipients (how many?)
>>>> -Somehow we monitor for disappearing members and then recast the leader
>>>> election if necessary.
>>>> 
>>>> Paxon protocol would be involved, I’d guess.  Does anyone have
>> references
>>>> to any academic work on presence monitoring and leader election, beyond
>>>> Lamport’s original paper?
>>>> 
>>>> I also wonder, is there a reason not to just use Multicast if it’s
>>>> available (I realize that it isn’t always supported - Amazon EC2, for
>>>> instance).
>>>> 
>>>> Interesting question!
>>>> 
>>>> Cheers,
>>>> 
>>>> Greg Trasuk
>>>> 
>>>>> On Jul 29, 2015, at 12:40 PM, Bryan Thompson <bryan@systap.com
>> <javascript:;>> wrote:
>>>>> 
>>>>> Hello,
>>>>> 
>>>>> I am wondering if anyone has looked into creating tree based algorithms
>>>> for
>>>>> multi-cast of RMI messages for river.  Assuming a local cluster, such
>>>>> patterns generally have log(p) cost for a cluster with p nodes.
>>>>> 
>>>>> For the curious, this is how many MPI messages are communicated under
>> the
>>>>> hood.
>>>>> 
>>>>> Thanks,
>>>>> Bryan
>>>> 
>>>> 
>> 
>> 
> 
> -- 
> ----
> Bryan Thompson
> Chief Scientist & Founder
> SYSTAP, LLC
> 4501 Tower Road
> Greensboro, NC 27410
> bryan@systap.com
> http://blazegraph.com
> http://blog.bigdata.com <http://bigdata.com>
> http://mapgraph.io
> 
> Blazegraph™ <http://www.blazegraph.com/> is our ultra high-performance
> graph database that supports both RDF/SPARQL and Tinkerpop/Blueprints
> APIs.  Blazegraph is now available with GPU acceleration using our disruptive
> technology to accelerate data-parallel graph analytics and graph query.
> 
> CONFIDENTIALITY NOTICE:  This email and its contents and attachments are
> for the sole use of the intended recipient(s) and are confidential or
> proprietary to SYSTAP. Any unauthorized review, use, disclosure,
> dissemination or copying of this email or its contents or attachments is
> prohibited. If you have received this communication in error, please notify
> the sender by reply email and permanently delete all copies of the email
> and its contents and attachments.


Mime
View raw message