Biosequence Similarity Search on the Mercury System

1
Biosequence Similarity Search on the Mercury
System

• Praveen Krishnamurthy, Jeremy Buhler, Roger
  Chamberlain, Mark Franklin, Kwame Gyang, and
  Joseph Lancaster
• Department of Computer Science and Engineering,
  Washington University in Saint Louis, MO

Supported by an NIH STTR Grant NSF Grants
DBI-0237902, ITR-0313203, CCR-0217334
2
Outline

• Overview of BLAST
• Overview of the Mercury system
• Description of BLASTN algorithm
• Algorithmic changes to BLASTN
• Improvement in performance
• Related work
• Conclusion

3
Basic Local Alignment Search Tool

• Biosequence comparison software
• Query sequence (new genome) to large database of
  known biosequences
• Look for similar regions
• Exponential growth of genomic databases
• Longer time for searches to complete
• Solutions
• Perform comparison over multiple machines
• Specialized hardware - Our Approach

4
The Mercury System
5
The Mercury System

• Proximity to disk
• Simple operations performed close to disk
• Avoids CPU use
• 400 Mbytes/s throughput from the disk
• Concurrent Independent operation
• Does not use processor cache cycles, memory or
  I/O buses
• Reconfigurable logic
• Logic can be tuned to the particular need of the
  application

6
BLASTN

• BLASTN
• Both the query and the database are long DNA
  strings
• Consist of A, C, T, G and some unknowns
• Each stage processes lesser data
• The stages become more computationally expensive

7
BLASTN - Terminology
Query
ACTGTGTTTCACTGACGGGTGT
Database
CTGTGTCCCCAACACTGCTGACGTAGAATCGTGTAG
w-mer is a sequence of w consecutive bases
8
BLASTN - Pipeline - Stage 1

• Matches each 11-mer in query to database
• Exact string matching
• 83 of overall time is spent in this stage
• Filters 92 of data entering this stage
• Only 8 of data proceeds to the next stage

9
BLASTN - Pipeline - Stage 2

• Extends the matches from stage 1

ACTGTGTTTCACTGACGGGTGT
GTGTCCCCAACATTTCACTGACGAGAATCGTGTAG
10
BLASTN - Pipeline - Stage 2

• Extends the matches from stage 1
• Allows mismatches of individual bases
• Does not allow gaps in either the query or the
  database
• Match score should be higher than threshold to
  proceed
• 16 of pipeline time is spent is this stage
• Only 2/100,000 of data entering this stage
  proceeds to the next stage

11
BLASTN - Pipeline - Stage 3

• Extends the matches from stage 2

ACCACTGTTTCACTGACG_GA_T_GT
CTGTGTCCCCAC_GTTTCACTGACGAGAATCGTGTAG
12
BLASTN - Pipeline - Stage 3

• Extends the matches from stage 2
• Scores matches with Gaps inserted in both the
  sequences
• Smith-Waterman dynamic programming algorithm
• lt1 of pipeline time is spent is this stage

13
NCBI - BLASTN

• Stage 1 (word matching) is implemented as a
  lookup table
• Efficient only for certain word lengths (w 11)
• Performance degrades dramatically for larger
  query sizes

Pentium-4 2.6GHz 1Gbyte RAM
14
Firmware implementation - Stage 1
Eliminates false-positives from Bloom filters,
obtain offset in query
Discards matches that are close to one another
Matches 11-mers to query, but generates
false-positives
15
Bloom filters operation
Programming the query into the bloom filter
(processing query)
K Hash Functions
query
11-mer
m-bit vector
16
Bloom filters operation
Finding matches in the database
1 Potential match
K Hash Functions
database
11-mer
0 Not a match
m-bit vector
17
Bloom filters operation
Finding matches in the database
?
1 Potential match
K Hash Functions
?
database
11-mer
0 Not a match
?
m-bit vector
False positives are eliminated using a hash
table
18
Bloom filter performance
19
Performance analysis
Firmware Vs. Software Stage 1
20
Overall system throughput
Tputoverall min (Tput1, Tput(23))
21
Stage 2 in firmware - Throughput
22
Stage 2 in firmware - Speedup
23
Related work

• Hardware based commercial systems
• Paracel GeneMatcherTM, used ASIC, and hence is
  inflexible
• RDisk, FPGA based system with throughput of 60
  Mbases/s for stage 1
• High-end commercial system
• Paracel BLASTMachine2TM, 32 CPU linux cluster
• 2.93 Mbases/s for 2.8 Mbase query
• 2 times faster than 1-node Mercury BLASTN
• TimeLogic DeCypherBLASTTM, FPGA based
• 213 Kbases/s for a 16 Mbase query
• Comparable to 1-node Mercury BLASTN

24
Conclusion

• BLASTN on the Mercury system
• Bloom filters to improve performance of stage 1
• Efficient hash functions in hardware
• 7x improvement in speed with only stage 1
  firmware
• gt50x speedup with stage 2 implemented in firmware
• Future work
• Algorithmic changes to stage 2
• Efficient use of hardware capabilities
• Other apps
• BLASTP, BLASTX etc.

25
Thank you