GraphLab

[Shannon Quinn <sq...@gatech.edu>]

Being the newbie on the block, forgive me if I'm rehashing old news: has 
anything seen/heard of GraphLab before?

http://www.graphlab.ml.cmu.edu/index.html

It's written by someone who has an office in the same exact building as 
I do, just one floor up, so I'll certainly be talking to him soon. But 
if there is someone here who is familiar with this work, can you 
elaborate on the differences between it and Mahout? He seems to have 
somewhat tweaked the standard map/reduce paradigm into something that 
offers more crosstalk flexibility between nodes at runtime (at the cost 
of significant configurational overhead, most likely), but beyond that 
it seems strikingly similar to the functionality Mahout provides.

Anyway, was pointed to this by someone in my department while I was 
running my coalescing thesis ideas by him.

Shannon

Re: GraphLab

[Lance Norskog <go...@gmail.com>]

Or the other way around: I've been thinking of the Mahout jobs as Pig
primitives.


On Tue, Mar 8, 2011 at 3:26 PM, Sean Owen <sr...@gmail.com> wrote:
> Looks interesting -- it looks like a specialization for iterative
> algorithms of a certain kind, a kind which describes a lot of
> algorithms. Is this distributed? It looked more like it's intended for
> high-performance machines. I guess it's also different being C++-based
> and not Hadoop-based.
>
> Hadoop is, in the end, a tool that was never conceived for general
> distributed computation. But among frameworks it's (relatively) well
> understood and available. It seems like Mahout has taken on the
> mission of delivering something that works on the framework that's out
> there now, which is a practical rather than theoretically-motivated
> goal. (I think it's a good goal too.) I see that as a difference from
> many research-oriented projects.
>
> Beyond that it is the same sort of thing and that's good.
>
> The thing I "worry" most that is being duplicated is actually Pig. It
> at least gives something more like "primitives" for basic
> information-shuffling operations on Hadoop like the sorts of pivots
> and joins and filters that go into your standard implementation of an
> ML algorithm. I bet we'd find we'd be better off bringing in some
> stuff from Pig rather than reinvent the join a few times over.
>
> But first things first... would really be good to focus on revamping
> and bringing together what we have already to pull together
> commonality and such before thinking what we can improve about those
> commonalities.
>
>
> On Tue, Mar 8, 2011 at 11:07 PM, Shannon Quinn <sq...@gatech.edu> wrote:
>> Being the newbie on the block, forgive me if I'm rehashing old news: has
>> anything seen/heard of GraphLab before?
>>
>> http://www.graphlab.ml.cmu.edu/index.html
>>
>> It's written by someone who has an office in the same exact building as I
>> do, just one floor up, so I'll certainly be talking to him soon. But if
>> there is someone here who is familiar with this work, can you elaborate on
>> the differences between it and Mahout? He seems to have somewhat tweaked the
>> standard map/reduce paradigm into something that offers more crosstalk
>> flexibility between nodes at runtime (at the cost of significant
>> configurational overhead, most likely), but beyond that it seems strikingly
>> similar to the functionality Mahout provides.
>>
>> Anyway, was pointed to this by someone in my department while I was running
>> my coalescing thesis ideas by him.
>>
>> Shannon
>>
>



-- 
Lance Norskog
goksron@gmail.com

Re: GraphLab

[Ted Dunning <te...@gmail.com>]

Including zookeeper is a major improvement. 

The new cluster executive does nothing to improve the lamentable state of hdfs. 

Sent from my iPhone

On Mar 8, 2011, at 6:06 PM, Dmitriy Lyubimov <dl...@gmail.com> wrote:

> PS one potential downside (or upside, perhaps) is that they are going
> to introduce
> zookeeper as a dependency into task management substrate. On the
> bright side, that would solve
> problem of namenode SPOF though (as far as i understand), among other things.

Re: GraphLab

[Dmitriy Lyubimov <dl...@gmail.com>]

PS one potential downside (or upside, perhaps) is that they are going
to introduce
zookeeper as a dependency into task management substrate. On the
bright side, that would solve
problem of namenode SPOF though (as far as i understand), among other things.

On Tue, Mar 8, 2011 at 6:04 PM, Dmitriy Lyubimov <dl...@gmail.com> wrote:
> On Tue, Mar 8, 2011 at 3:26 PM, Sean Owen <sr...@gmail.com> wrote:
>> Looks interesting -- it looks like a specialization for iterative
>
>> Hadoop is, in the end, a tool that was never conceived for general
>> distributed computation. But among frameworks it's (relatively) well
>> understood and available. It seems like Mahout has taken on the
>> mission of delivering something that works on the framework that's out
>> there now, which is a practical rather than theoretically-motivated
>> goal. (I think it's a good goal too.) I see that as a difference from
>> many research-oriented projects.
>>
>
> At the last HUG they rolled out plans (preliminary alpha ETA summer) where
> they separate task management substrate from application substrate. I.e. once
> you got your task allocation & data/rack affinity refactored as a
> standalone concern,
> you can run MR or even MPI or whatever distributed data flow your
> heart desires.
>
> That's IMO a good news for stuff like mahout-math, a lot of times
> matrix jobs require something
> that is currently emulated by map-only passes, or has to resort to
> reduction whereas all is though
> is sequential merge without sort component .
>
> So i think brighter days are ahead (for Mahout in particular).
>

Re: GraphLab

[Dmitriy Lyubimov <dl...@gmail.com>]

On Tue, Mar 8, 2011 at 3:26 PM, Sean Owen <sr...@gmail.com> wrote:
> Looks interesting -- it looks like a specialization for iterative

> Hadoop is, in the end, a tool that was never conceived for general
> distributed computation. But among frameworks it's (relatively) well
> understood and available. It seems like Mahout has taken on the
> mission of delivering something that works on the framework that's out
> there now, which is a practical rather than theoretically-motivated
> goal. (I think it's a good goal too.) I see that as a difference from
> many research-oriented projects.
>

At the last HUG they rolled out plans (preliminary alpha ETA summer) where
they separate task management substrate from application substrate. I.e. once
you got your task allocation & data/rack affinity refactored as a
standalone concern,
you can run MR or even MPI or whatever distributed data flow your
heart desires.

That's IMO a good news for stuff like mahout-math, a lot of times
matrix jobs require something
that is currently emulated by map-only passes, or has to resort to
reduction whereas all is though
is sequential merge without sort component .

So i think brighter days are ahead (for Mahout in particular).

Re: GraphLab

[Sean Owen <sr...@gmail.com>]

Looks interesting -- it looks like a specialization for iterative
algorithms of a certain kind, a kind which describes a lot of
algorithms. Is this distributed? It looked more like it's intended for
high-performance machines. I guess it's also different being C++-based
and not Hadoop-based.

Hadoop is, in the end, a tool that was never conceived for general
distributed computation. But among frameworks it's (relatively) well
understood and available. It seems like Mahout has taken on the
mission of delivering something that works on the framework that's out
there now, which is a practical rather than theoretically-motivated
goal. (I think it's a good goal too.) I see that as a difference from
many research-oriented projects.

Beyond that it is the same sort of thing and that's good.

The thing I "worry" most that is being duplicated is actually Pig. It
at least gives something more like "primitives" for basic
information-shuffling operations on Hadoop like the sorts of pivots
and joins and filters that go into your standard implementation of an
ML algorithm. I bet we'd find we'd be better off bringing in some
stuff from Pig rather than reinvent the join a few times over.

But first things first... would really be good to focus on revamping
and bringing together what we have already to pull together
commonality and such before thinking what we can improve about those
commonalities.


On Tue, Mar 8, 2011 at 11:07 PM, Shannon Quinn <sq...@gatech.edu> wrote:
> Being the newbie on the block, forgive me if I'm rehashing old news: has
> anything seen/heard of GraphLab before?
>
> http://www.graphlab.ml.cmu.edu/index.html
>
> It's written by someone who has an office in the same exact building as I
> do, just one floor up, so I'll certainly be talking to him soon. But if
> there is someone here who is familiar with this work, can you elaborate on
> the differences between it and Mahout? He seems to have somewhat tweaked the
> standard map/reduce paradigm into something that offers more crosstalk
> flexibility between nodes at runtime (at the cost of significant
> configurational overhead, most likely), but beyond that it seems strikingly
> similar to the functionality Mahout provides.
>
> Anyway, was pointed to this by someone in my department while I was running
> my coalescing thesis ideas by him.
>
> Shannon
>