Month: March 2013

by

Hadoop and MPI

At this point of writing, I am reading the Hadoop the definitive Guide, 3rd Edition from Oreilly. I thought I capture some information down from what I learn.

Here is a summary of the difference between MPI and MapReduce

MPI MapReduce
Location of Data
 Shared Storage Data Locality (within the Node)
Complexity MPI require the programmer to handle
the mechanics of the data flow exposed
via low-level C routines and constructs		 Programmer in terms of functions
of key and value pairs, and the data flow
is implicit.
Compute
Characteristics
 Large-Scale Distributed Computation.
A failed process will slow or halt the
progress of the computation. MPI has to
explicitly manage checkpoint and recovery		 Shared Nothing Architecture. The implementation
detects  failed map or reduce tasks and
reschedules  replacements on machines that
are healthy
by

Hadoop and Traditional RDMS

At this point of writing, I am reading the Hadoop the definitive Guide, 3rd Edition from Oreilly. I thought I capture some information down from what I learn.

Here is a summary of the difference between Traditional RDMS and MapReduce

Traditional RDBMS MapReduce
Data size Gigabytes Petabytes
Access Interactive and batch Batch
Updates Read and write many times Write once, read many times
Structure Static schema Dynamic schema
Integrity High Low
Scaling Nonlinear Linear
by

PBS scripts for mpirun parameters for Chelsio / Infiniband Cards

If you are running Chelsio Cards, you  may want to specify the mpirun parameters to ensure the

/usr/mpi/intel/openmpi-1.4.3/bin/mpirun 
-mca btl openib,sm,self --bind-to-core 
--report-bindings -np $NCPUS -machinefile $PBS_NODEFILE $PBS_O_WORKDIR/$file

–bind-to-core: Bind each MPI process to a core
–mca btl openib,sm,self: (Infiniband, shared memory, the loopback)

For information on Interprocess communication with shared memory,

  1. see Speaking UNIX: Interprocess communication with shared memory
by

Sequential execution of the Parallel or serial jobs on OpenPBS / Torque

If you have an requirement to execute jobs in sequence after the 1st job has completed, only then the 2nd job can launch, you can  use the -W command

For example, if you have a running job with a job ID 12345, and you run the next job to run only after job 12345 run.

$ qsub -q clusterqueue -l nodes=1:ppn=8 -W depend=afterany:12345 parallel.sh -v file=mybinaryfile

You will notice that the job will hold until the 1st job executed.

.....
.....
24328 kittycool Hold 2 10:00:00:00 Sat Mar 2 02:20:12