performance optimizations for GenotypeGVCFs

[akiezun]

part of work tracked here #1608

@lbergelson please review. The corresponding protected branch is https://github.com/broadinstitute/gatk-protected/tree/ak_ggvcf_performance

[coveralls]

Coverage increased (+0.008%) to 81.686% when pulling 321c561 on ak_ggvcf_performance into d0db982 on master.

[lbergelson]

If totalK is 0 (or less) this produces a different result than before. I'm not sure if that's possible or not.

[akiezun]

good catch. i think totalK may be 0 sometimes. I'll leave this as for ( int j = 1, n = targetSetLog10Likelihoods.length; j < n; j++ ) {

[lbergelson]

@akiezun minor comments

[coveralls]

Coverage increased (+0.0008%) to 81.699% when pulling 3d63672 on ak_ggvcf_performance into 1637de1 on master.

[akiezun]

back to @lbergelson for second review. log10 cache is now static (and synchronized at expansion only) as in gatk3 - it looks faster (no map lookup that would be needed in the thread local solution)

[lbergelson]

@akiezun My only concern now is that someone takes log10 of a very large number triggering a massive and slow cache expansion. This caching scheme is good for clustered queries of small values, but terrible for sparse large queries. Is that a case we need to consider?

[akiezun]

The access pattern I saw is that the queries start small and creep up to
tens or hundreds of thousands (Order of the queried numbers). With the
exponential cache expansion, the cache quickly catches up. So it's fine.
We'll revise when needed.

On Wednesday, July 6, 2016, Louis Bergelson notifications@github.com
wrote:

@akiezun https://github.com/akiezun My only concern now is that someone
takes log10 of a very large number triggering a massive and slow cache
expansion. This caching scheme is good for clustered queries of small
values, but terrible for sparse large queries. Is that a case we need to
consider?

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#1957 (comment),
or mute the thread
https://github.com/notifications/unsubscribe/AB5rL32j0zFsPo1VQ4T6gxnEHvjpYZ02ks5qS_VhgaJpZM4JCXE9
.

Sent from Gmail Mobile

[lbergelson]

👍