“base pairs” (see http://en.wikipedia.org/wiki/Base_pair). Typical
sizes of various genomes are given in (http://en.wikipedia.org/wiki/Genome).
String matching can be used to find particular sequences in a
genome. Several string matching algorithms are described
Consider a program to find to find if a particular sequence of
base pairs is found in a genome, and if so, where and how
Your program will run on a cluster with
the following properties:
Number of nodes – 20
Number of processors per node 16 2.6 GHz Xeon
Memory per node 16 GB
GPU – 2 (NVIDIA CUDA) per node, 1024 stream processors and 4GB RAM, running at 1.5 GHz
local drives 1 T SATA , 6 GB/sec
NFS drive 10TB RAID, bandwidth limited by network
Switched Ethernet network
Latency L = 20 microseconds
Bandwidth B = 1Gb/sec == 100 Mbytes/sec for messages
larger than 32Kbytes
You may not need all the above information. If you feel you need some other system
property, feel free to assume some reasonable value (Try Wikipedia)
Assume the genome you are exploring and the sequence you are
trying to find, are both initially files on the NFS disk.
1. Parallel String Match algorithm – in MPI, OpenMP, CUDA or some
combination of these. Description in English and/or pseudocode is
sufficient. Is yoyur algorithm data parallel, task parallel or both?
Describe data transfer during computation (disk to program, process to process,
CPU – GPU and node – node). Describe how data is partitioned between processes,
shared between processes, or replicated at each process.
2. You may not need all the hardware available for your algorithm. You may use the
entire cluster or any part of it. Describe what resources your algorithm will use to
execute. Explain your choice.
3. Estimate how your algorithm would perform on the computer
system described above. Consider:
a. Complexity; communication costs.
b. Is there some file size (in bytes, number of elements, or both)
that is too small for your algorithm to work efficiently? Given the wide range of genome
sizes (see http://en.wikipedia.org/wiki/Genome), is there some range of size that you expect would
be best for your algorithm?
c. How much speedup would you exepect on the given hardware as compared to running
on a single CPU? Justify your answer