Merged
Conversation
CHANGELOG: Fixes hail-is#13704, in which Hail could encounter an IllegalArgumentException if there are too many transient errors. I need to do the multiplication in 64-bits so that it does not wrap around to a large negative value. Then I can use `math.min` with the maxDelayMs to get us back into 32-bits.
Contributor
Author
|
@ehigham bump, this is blocking Wenhan |
chrisvittal
approved these changes
Sep 27, 2023
danking
pushed a commit
to danking/hail
that referenced
this pull request
Oct 4, 2023
CHANGELOG: Reduce latency on simple pipelines by as much as 50% by reducing decoding time. Force count essentially tests decoding because it forces decoding but then just increments a counter by one. Analysis of profile results indicates that the array inplace decoder was perhaps 50% of time, but exactly what part of decoding was unclear. I attempted many different things. I eventually settled on loop unrolling as the primary benefit. After team meeting, I applied @patrick-schultz 's advice to use bit twiddling to further improve the speed. --- I assessed the latency using `time python3` on this file: ```python import hail as hl hl.init(master='local[1]') hl._set_flags(write_ir_files='1') hl.read_matrix_table('/Users/dking/projects/hail-data/foo.mt')._force_count_rows() ``` `foo.mt` is a subset of the `variant_data` from a VDS with ~80k samples, ~300k variants, stored in ~1.6GiB. 1. This PR: 34s, 33s 2. no twiddling: 43s, 43s hail-is/hail@main...danking:hail:unroll-64 3. no twiddling & 8 element blocks: 37s, 38s hail-is/hail@main...danking:hail:unroll-8 4. `main` (`481cfc201b [query] fix backoff code (hail-is#13713)`): 68s, 69s In YourKit, I observe that (1) reads 50-70MB/s with one core whereas (4) reads 15-35MB/s. I also assessed the 10-core latency and JIT effects: - (1) starts at ~12s, warms to ~6s (+- 0.5s). Peak bandwidth 490MB/s. - (4) starts at ~17s and warms up to ~11s (+- 2s). Peak bandwidth ~250MB/s. I suspect, with this PR, the multi-core speed is fast enough to saturate any of our file stores (including my laptop, which I think taps out just around ~500MB/s). Big thanks to everyone who contributed, particularly @patrick-schultz, whose suggestion to use bit-twiddling, squeezeed another 10% off the 8 element blocks.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
CHANGELOG: Fixes #13704, in which Hail could encounter an IllegalArgumentException if there are too many transient errors.
I need to do the multiplication in 64-bits so that it does not wrap around to a large negative value. Then I can use
math.minwith the maxDelayMs to get us back into 32-bits.I'm just pushing through a bunch of bugs to get Wenhan unblocked today.