Yes you are right, you should not replace grep for one off regexes due to the latency of the io operations. As far as I recall, memory throughput however is much higher on house than on CPUs (or at least this was true in 2012). Our intended use cases for this sort of a grep were cases where you are finding many needles in enormous haystacks such as: Looking for malicious machine code in executables that you download off of the internet (virus scanning), or searching for certain genetic sequences in DNA code. I believe we discussed this in our final presentation, but perhaps we left it out of the written report.
Yes we probably should have included the numbers to show why you wouldn't want to use this instead of grep for simple tasks. I am sorry we missed it.
(Also, you'd want BLAST for DNA instead of grep, but perhaps this was an intentional simplification)
My last K-mer counter I wrote was CPU/Memory bound (big enough size strings means big hash tables to store this stuff in, or using something like a bloom filter (Jellyfish's approach), not IO bound.
Anyway, for anyone who wants to write something fun, fasta parsers / k-mer counters are a good weekend project.
here was my work in the area (shameless plug, though i'm not in the field anymore)
https://github.com/mutantturkey/dna-utils.
just tested again against a 400mbish genome fasta file from ncbi and it's still clearly an cpu/memory issue with huge hash tables. See below (4^x is the potential combinations of K-mers). K=9 is done with a 4^9 array (sane in memory), 4^12 is a sparse array (stored as a a std::unordered_map, because it's too large to allocate that much space in memory). One is obviously way slower
calvin@bison:~/src/dna-utils$ time ./kmer_total_count -k 9 -i 3816_ref_Abrus_2018_chrUn.fa.2 >/dev/null
real 0m4.235s
user 0m4.116s
sys 0m0.108s
calvin@bison:~/src/dna-utils$ time ./kmer_total_count -k 12 -i 3816_ref_Abrus_2018_chrUn.fa.2 >/dev/null
real 0m25.524s
user 0m25.292s
sys 0m0.152s
calvin@bison:~/src/dna-utils
https://www.microsoft.com/en-us/microsoft-365/windows/micros...
[0] https://www.dfrws.org/sites/default/files/session-files/pres...
If we are using something more 'grep'-like, it's for large quantities of data, and in many cases we could treat these as a stream. So as long as the bandwidth to the GPU is sufficient, a little stall at the beginning would represent an undue hardship, if the results come back faster.
Also, they didn't ignore the key detail, they mention it: "This is without taking into account the overhead of transferring data to the GPU."
Fair point, it's more support for this HN post[2] making the case for memory-mapping files to GPU.
> This is a neat project, but there is a reason we are not using the GPU for regex matching.
I think it's more interesting that it's a general facility for running automata on a GPU, as it's opening up the universe of things that can be accelerated by a GPU. For instance, Numba[1] is a library to run arbitrary math code, and it has a feature for generating code to run on the GPU.
[1]: http://numba.pydata.org/numba-doc/latest/cuda/index.html
Also, IO bound? Again, no. My consumer MBP can read at 3.2GB/s. Good luck doing regex on that in real time on a CPU. This is a CPU bound operation for a typical high end configuration.
Does this make sense for a 1MB file? probably not. But anything well beyond that it will be way faster.
Here's a great talk about how you should think about GPU performance: https://www.nvidia.com/content/GTC-2010/pdfs/2238_GTC2010.pd...
To give some context: PCIe 3.0 x16 achieves about 11-12 GB/s. So, transfers for benchmarked 53MB file would add ~4.8 ms (per direction, assuming pinned host memory)
Now, they could also pursue a streaming approach, processing in "micro-batches". I.e., pipelining input transfers (HOST->GPU), GPU-based processing, and transfer of results (GPU->HOST), as pursued by [1] (disclaimer: I'm a co-author). This lower's processing latency. Since the interconnect is full-duplex, it would just limit the processing rate to about ~11 GB/s (i.e., 5 ms for mentioned 53MB). To be fair, though, input has to be sufficiently large, such that there are (a) enough micro-batches to overlap transfer and compute and (b) the micro-batches can be chosen to be still large enough to exploit full parallelism of the GPU.