Share
Embed
Awesome Big Data Algorithms by Titus Brown
Random algorithms and probabilistic data structures are algorithmically efficient and can provide shockingly good practical results. I will give a practical introduction, with live demos and bad jokes, to this fascinating algorithmic niche. I will conclude with some discussions of how our group has applied this to large sequencing data sets (although this will not be the focus of the talk).
PyCon 2013
March 15, 2013
Transcript
37 slides
-
Awesome Big Data Algorithms C. Titus Brown
[email protected] Asst Professor, Michigan State University (Microbiology, Computer Science, and BEACON) -
Welcome! • More of a computational scientist than a
computer scientist; will be using simulations to demo & explore algorithm behavior. • Send me questions/comments @ctitusbrown, or [email protected]. -
“Features” • I will be using Python rather than
C++, because Python is easier to read. • I will be using IPython Notebook to demo. • I apologize in advance for not covering your favorite data structure or algorithm. -
Outline • The basic idea • Three examples
– Skip lists (a fast key/value store) – HyperLogLog Counting (counting discrete elements) – Bloom filters and CountMin Sketches • Folding, spindling, and mutilating DNA sequence • References and further reading -
The basic idea • Problem: you have a lot
of data to count, track, or otherwise analyze. • This data is Data of Unusual Size, i.e. you can’t just brute force the analysis. • For example, – Count the approximate number of distinct elements in a very large (infinite?) data set – Optimize queries by using an efficient but approximate prefilter – Determine the frequency distribution of distinct elements in a very large data set. -
Online and streaming vs. offline “Large is hard; infinite
is much easier.” • Offline algorithms analyze an entire data set all at once. • Online algorithms analyze data serially, one piece at a time. • Streaming algorithms are online algorithms that can be used for very memory & compute limited analysis. -
Exact vs random or probabilistic • Often an approximate
answer is sufficient, esp if you can place bounds on how wrong the approximation is likely to be. • Often random algorithms or probabilistic data structures can be found with good typical behavior but bad worst case behavior. -
For one (stupid) example You can trim 8 bits
off of integers for the purpose of averaging them -
Skip lists A randomly indexed improvement on linked lists.
Each node can belong to one or more vertical “levels”, which allow fast search/insertion/deletion – ~O(log(n)) typically! wikipedia -
Skip lists A randomly indexed improvement on linked lists.
Very easy to implement; asymptotically good behavior. From reddit, “if someone held a gun to my head and asked me to implement an efficient set/map storage, I would implement a skip list.” (Response: “does this happen to you a lot??”) wikipedia -
Channel randomness! • If you can construct or rely
on randomness, then you can easily get good typical behavior. • Note, a good hash function is essentially the same as a good random number generator… -
HyperLogLog cardinality counting • Suppose you have an incoming
stream of many, many “objects”. • And you want to track how many distinct items there are, and you want to accumulate the count of distinct objects over time. -
Relevant digression: • Flip some unknown number of coins.
Q: what is something simple to track that will tell you roughly how many coins you’ve flipped? • A: longest run of HEADs. Long runs are very rare and are correlated with how many coins you’ve flipped. -
Cardinality counting with HyperLogLog • Essentially, use longest run
of 0-‐bits observed in a hash value. • Use multiple hash functions so that you can take the average. • Take harmonic mean + low/high sampling adjustment => result. -
Bloom filters • A set membership data structure that
is probabilistic but only yields false positives. • Trivial to implement; hash function is main cost; extremely memory efficient. -
My research applications Biology is fast becoming a data-‐driven
science. http://www.genome.gov/sequencingcosts/ -
Shotgun sequencing analogy: feeding books into a paper shredder,
digitizing the shreds, and reconstructing the book. Although for books, we often know the language and not just the alphabet J -
Shotgun sequencing is -‐-‐ • Randomly ordered. •
Randomly sampled. • Too big to efficiently do multiple passes -
Shotgun sequencing Genome (unknown) X X X X X
X X X X X X X X X Reads (randomly chosen; have errors) X X X “Coverage” is simply the average number of reads that overlap each true base in genome. Here, the coverage is ~10 – just draw a line straight down from the top through all of the reads. -
Random sampling => deep sampling needed Typically 10-‐100x needed
for robust recovery (300 Gbp for human) -
Random sampling => deep sampling needed Typically 10-‐100x needed
for robust recovery (300 Gbp for human) But this data is massively redundant!! Only need 5x systematic! All the stuff above the red line is unnecessary! -
Streaming algorithm to do so: digital normalization True
sequence (unknown) Reads (randomly sequenced) -
Digital normalization True sequence (unknown) Reads (randomly sequenced) X
X X X X X X X X X X -
Digital normalization True sequence (unknown) Reads (randomly sequenced) X
X X X X X X X X X X -
Digital normalization True sequence (unknown) Reads (randomly sequenced) X
X X X X X X X X If next read is from a high coverage region - discard X X -
Digital normalization True sequence (unknown) Reads (randomly sequenced) X
X X X X X X X X X X X X X X X X X X X X X X X Redundant reads (not needed for assembly) -
Storing data this way is better than best-‐ possible information-‐theoretic
storage. Pell et al., PNAS 2012 -
Use Bloom filter to store graphs Pell et al.,
PNAS 2012 Graphs only gain nodes because of Bloom filter false positives. -
Some assembly details • This was completely intractable.
• Implemented in C++ and Python; “good practice” (?) • We’ve changed scaling behavior from data to information. • Practical scaling for ~soil metagenomics is 10x: – need < 1 TB of RAM for ~2 TB of data, ~2 weeks. – Before, ~10TB. • Smaller problems are pretty much solved. • Just beginning to explore threading, multicore, etc. (BIG DATA grant proposal) • Goal is to scale to 50 Tbp of data (~5-‐50 TB RAM currently) -
Concluding thoughts • Channel randomness. • Embrace streaming.
• Live with minor uncertainty. • Don’t be afraid to discard data. (Also, I’m an open source hacker who can confer PhDs, in exchange for long years of low pay living in Michigan. E-‐mail me! And don’t talk to Brett Cannon about PhDs first.) -
References SkipLists: Wikipedia, and John Shipman’s code: http://infohost.nmt.edu/tcc/help/lang/python/examples/pyskip/pyskip.pdf
HyperLogLog: Aggregate Knowledge’s blog, http://blog.aggregateknowledge.com/2012/10/25/sketch-‐of-‐the-‐day-‐hyperloglog-‐cornerstone-‐of-‐a-‐big-‐data-‐infrastructure/ And: https://github.com/svpcom/hyperloglog Bloom Filters: Wikipedia Our work: http://ivory.idyll.org/blog/ and http://ged.msu.edu/interests.html [email protected]