Use a clock-based Buffer Manager eviction strategy #3620
Merged
Conversation
This file contains 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
benjaminwinger
force-pushed
the
eviction-queue-opt
branch
from
07fcd85
to
66a7800
Compare
benjaminwinger
changed the title
|
Can u also benchmark on some pure scan queries on large dataset like ldbc100 comment table? I think the cold run should be able to present performance difference. |
Benchmark ResultMaster commit hash:
|
ray6080
approved these changes
Jun 11, 2024
There was a problem hiding this comment.
Nice PR!
One thing to discuss here: I feel we should start to take the mem usage of the queue into consideration. BM's available memory should be bufferPoolSize - queueSize. Though the ratio of queue's memory usage is small, the principle is to try to make memory usage more accurate when possible and convenient. What do you think?
I think that's reasonable. |
There is a difference, but it doesn't appear to be huge.
(both run cold from a database stored on an ssd). |
benjaminwinger
force-pushed
the
eviction-queue-opt
branch
from
66a7800
to
8e455e4
Compare
Benchmark ResultMaster commit hash:
|
benjaminwinger
force-pushed
the
eviction-queue-opt
branch
2 times, most recently
from
b6bbea8
to
77a4a31
Compare
Instead of queueing potential pages to evict when unpinning, all pages currently in memory are kept in the evictionqueue. When more space is needed, pages are evicted if they haven't changed since the last time we scanned them. Based on the hash table implementation in the VMCache paper, except that we store page states by reference and for simplicity store the eviction candidates in a circular buffer instead of a hash table since lookups aren't necessary.
benjaminwinger
force-pushed
the
eviction-queue-opt
branch
from
77a4a31
to
0631ce8
Compare
Benchmark ResultMaster commit hash:
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Instead of queueing potential pages to evict when unpinning, all pages currently in memory are kept in the evictionqueue. When more space is needed, pages are evicted if they haven't changed since the last time we scanned them. The eviction queue is now only accessed when fetching pages from disk and has no global lock.
Fixes #2137
Based on the hash table implementation in the VMCache paper, except that we store page states by reference and for simplicity store the eviction candidates in a circular buffer instead of a hash table since lookups aren't necessary.
While this isn't the first place we're using atomics, GCC at least depends on libatomic for 16-byte compare and swap because it's not generating the 16-byte compare exchange instruction (cmpxchg16b) directly (see the open bug, and also the atomic docs). Even with libatomic, it won't use cmpxchg16b without custom flags and instead uses a fallback (the
-mcx16flag is for this specifically, but-march=nativewill also enable it if your CPU supports it or-march=x86_64-v2or newer). Benchmarks with-mcx16were almost identical, so I don't think this is much of a bottleneck (using 8-byte atomic eviction and insertion cursors should make it rare that two different threads are trying to evict or insert to the same spot anyway).Performance improvements are minimal when there is plenty of buffer room.
I tried running some large copy benchmarks (two copies of 60M integers) with a restricted buffer, and found that the master version kept throwing buffer manager exceptions if I didn't give it a buffer pool of at least 1.5GB. With this version it works with the minimum buffer pool size of 64MB, and there was also a significant performance improvement when running this version with a 1.5GB buffer pool compared to the original.
Benchmarks are fairly rough; I ran each at least twice and averaged them (times are the first/second copies, each of 60M integers).