Introduction of the PrefetchRateLimiter #13907
Conversation
| @@ -1345,6 +1410,14 @@ void BlockBasedTableIterator::FindBlockForwardInMultiScan() { | |||
| } | |||
| // Move to the next pinned data block | |||
| ResetDataIter(); | |||
| if (multi_scan_->prefetch_rate_limiter) { | |||
| size_t releasing = | |||
| multi_scan_->pinned_data_blocks[multi_scan_->cur_data_block_idx] | |||
There was a problem hiding this comment.
pinned_data_blocks[multi_scan_->cur_data_block_idx] would not be valid at this point I think, since it would've been transferred to the data block iter. Or am I missing something?
| virtual bool release(size_t bytes) = 0; | ||
| }; | ||
|
|
||
| class DefaultPrefetchRateLimiter : public PrefetchRateLimiter { |
There was a problem hiding this comment.
This need not be declared in a public header file. RocksDB typically just exposes an allocator, like NewDefaultPrefetchRateLimiter().
| multi_scan_->pinned_data_blocks[multi_scan_->cur_data_block_idx] | ||
| .GetValue() | ||
| ->size(); | ||
| multi_scan_->prefetch_rate_limiter->release(releasing); |
There was a problem hiding this comment.
Ideally we'd do this in a cleanup function registered with block_iter_ (which is derived from Cleanable) so that the release happens whenever block_iter_ is reset.
| // Lets make sure we are rate limited on how many blocks to prepare | ||
| if (multiscan_opts->prefetch_rate_limiter) { | ||
| auto blocks = multiscan_opts->GetMutablePrefetchRateLimiter().acquire( | ||
| table_, index_iter_->value().handle.size(), true); |
There was a problem hiding this comment.
Write last parameter as /*all_or_nothing=*/true (Google C++ style guide - https://google.github.io/styleguide/cppguide.html#Function_Argument_Comments)
|
Its not clear how the proposed interface would support prefetching across multiple levels. For example, say we have L1, L2 and L3. Suppose L1's Prepare() gets called first. It could potentially exhaust the prefetch quota and L2/L3 cannot do any prefetching. I understand that we want to keep the initial implementation simple for now that works for single level use cases, but the interface should be extensible to support other use cases in the future. |
The multiscan operation uses the prepare function prefetch and pin all the required blocks for its provided scan ranges. However, this could cause large amount of memory being used. To restrict this, we introduce the PrefetchRateLimiter which allows users to bound how much is prefetched to ensure we do not OOM the host.
Test Plan
Added additional tests.