On Jan 11, 2012, at 10:15 AM, George Kousiouris wrote: > > Hi, > > see comments in text > > On 1/11/2012 4:42 PM, Merto Mertek wrote: >> Hi, >> >> I was wondering if anyone knows any paper discussing and comparing the >> mentioned topic. I am a little bit confused about the classification of >> hadoop.. Is it a /cluster/comp grid/ a mix of them? > I think that a strict definition would be an implementation of the map-reduce > computing paradigm, for cluster usage. > >> What is hadoop in >> relation with a cloud - probably just a technology that enables cloud >> services.. > It can be used to enable cloud services through a service oriented framework, > like we are doing in > http://users.ntua.gr/gkousiou/publications/PID2095917.pdf > > in which we are trying to create a cloud service that offers MapReduce > clusters as a service and distributed storage (through HDFS). > But this is not the primary usage. This is the back end heavy processing in a > cluster-like manner, specifically for parallel jobs that follow the MR logic. > >> >> Can it be compared to cluster middleware like beowulf, oscar, condor, >> sector/sphere, hpcc, dryad, etc? Why not? > I could see some similarities with condor, mainly in the job submission > processes, however i am not really sure how condor deals with parallel jobs. >
Since you asked…
<condor-geek>
Condor has a built-in concept of a set of jobs (called a "job cluster"). On
top of its scheduler, there is a product called "DAGMan" (DAG = directed
acyclic graph) that can manage a large number of jobs with interrelated
dependencies (providing a partial ordering between jobs). Condor with DAG is
somewhat comparable to the concept of Hadoop tasks plus Oozie workflows
(although the data aspects are very different - don't try to stretch it too
far).
Condor / PBS / LSF / {OGE,SGE,GE} / SLURM provide the capability to start many
identical jobs in parallel for MPI-type computations, but I consider MPI wildly
different than the sort of workflows you see with MapReduce. Specifically,
"classic MPI" programming (the ones you see in wide use, MPI2 and later are
improved) mostly requires all processes to start simultaneously and the job
crashes if one process dies. I think this is why the Top10 computers tend to
measure mean time between failure in tens of hours.
Unlike Hadoop, Condor jobs can flow between pools (they call this "flocking")
and pools can naturally cover multiple data centers. The largest demonstration
I'm aware of is 100,000 cores across the US; the largest production pool I'm
aware of is about 20-30k cores across 100 universities/labs on multiple
continents. This is not a criticism of Hadoop - Condor doesn't really have the
same level of data-integration as Hadoop does, so tackles a much simpler
problem (i.e., bring-your-own-data-management!).
</condor-geek>
Brian
smime.p7s
Description: S/MIME cryptographic signature
