Mas i370 patch d #383
Merged
Mas i370 patch d #383
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
In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes.
Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4
Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information.
This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt
erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes
As it is persisted as part of the file write, any change to the summary record cannot be rolled back
Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way
If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads.
To more consistently trigger all overload scenarios
Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue
- Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used
Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function.
There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs.
The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null.
|
Extend Patch C to ensure that both an excess bookie ledger cache, and an over-size penciller in-memory cache will be persisted to disk even where no traffic is present |
Closed
Unit test had a typo - and result interpretation had a misunderstanding.
Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place.
There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message.
|
This compares different configurations on a high-key-volume load test (one using small objects and lots of PUTs to try and rapidly increase the volume of keys). Memory -
Throughput -
It seems that we get improvements in memory efficiency almost for free in terms of throughput. The Profile A refers to setting eheap SBCT to 128KB (default 512KB): erlang.eheap_memory.sbct = 128 |
|
The memory is as reported by the OS, averaged across the nodes in the cluster. The test duration was 36 hours |
This works functionally, but is not optimised (the term is replicated in the index)
If the summary index all have the same index term - only the object keys need to be indexes
|
Tests run to include the latest version of PatchD, and also an additional control on OTP 25 (without leveled fixes):
The positive news is increased memory efficiency in OTP 25. However, this doesn't deliver the same level of improvements as the patch |
We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null
If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok.
When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys
Add assertion to confirm worthwhile optimisation
|
Further test. This starts from a cluster pre-populated with 360M objects, and then aggressively inserts more over 24 hours (with load shifting to be more read-based in the last hour). Memory footprint:
Throughput achieved:
The PatchD performs well in terms of both throughput and memory optimisation. Adding the alternative configuration profile (sbct = 128KB ), reduces throughput at the start of the tense - but consistently delivers a lower memory footprint. A longer test needs to be run from the same base, to understand the longer-term behaviour when load goes primarily GETs not PUTs. |
yorkshire-steve
approved these changes
Dec 5, 2022
martinsumner
added a commit
that referenced
this pull request
Dec 18, 2022
* Refactor penciller memory
In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen:
TimeBucket PatchA
a.0ms_to_1ms 18554
b.1ms_to_2ms 51778
c.2ms_to_3ms 696
d.3ms_to_5ms 220
e.5ms_to_8ms 59
f.8ms_to_13ms 40
g.13ms_to_21ms 364
h.21ms_to_34ms 277
i.34ms_to_55ms 34
j.55ms_to_89ms 17
k.89ms_to_144ms 21
l.144ms_to_233ms 31
m.233ms_to_377ms 45
n.377ms_to_610ms 52
o.610ms_to_987ms 59
p.987ms_to_1597ms 55
q.1597ms_to_2684ms 54
r.2684ms_to_4281ms 29
s.4281ms_to_6965ms 7
t.6295ms_to_11246ms 1
It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal.
The refactoring here is two-fold:
- tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs;
- change the update to the index to be a simple extension of a list, rather than any conversion.
This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes.
* Compress SST index
Reduces the size of the leveled_sst index with two changes:
1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare.
2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4
* Immediate hibernate
Reasons for delay in hibernate were not clear.
Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information.
* Refactor BIC
This patch avoids the following:
- repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET
- Stops re-reading of all elements to discover high modified date
Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt
* Use correct size in test results
erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes
* Don't change summary record
As it is persisted as part of the file write, any change to the summary record cannot be rolled back
* Clerk to prompt L0 write
Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes.
The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way
* Add push on journal compact
If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again.
This is only relevant to riak users with very off/full batch type workloads.
* Extend tests
To more consistently trigger all overload scenarios
* Fix range keys smaller than prefix
Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller.
Unit tests and ct tests added to expose the potential issue
* Tidy-up
- Remove penciller logs which are no longer called
- Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog
- Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline
- Alight testutil slow offer with standard slow offer used
* Tidy-up
Remove pre-otp20 references.
Reinstate the check that the starting pid is still active, this was added to tidy up shutdown.
Resolve failure to run on otp20 due to `-if` sttaement
* Tidy up
Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function.
There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs.
* Remove R16 relic
* Further testing another issue
The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey.
If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null.
* Fix unit test
Unit test had a typo - and result interpretation had a misunderstanding.
* Code and spec tidy
Also look to the cover the situation when the FirstKey is the same as the Prefix with tests.
This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place.
* Hibernate on BIC complete
There are three situations when the BIC becomes complete:
- In a file created as part of a merge the BIS is learned in the merge
- After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless...
- Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests.
In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data.
Previously on the the first base was covered. Now all three are covered through the bic_complete message.
* Test all index keys have same term
This works functionally, but is not optimised (the term is replicated in the index)
* Summaries with same index term
If the summary index all have the same index term - only the object keys need to be indexes
* Simplify case statements
We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null
* OK for M == N
If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix.
If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key).
The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok.
* Simplify
Correct the test to use a binary field in the range.
To avoid further issue, only apply filter when everything is a binary() type.
* Add test for head_only mode
When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys
* Revert previous change - must support typed buckets
Add assertion to confirm worthwhile optimisation
martinsumner
added a commit
that referenced
this pull request
Dec 18, 2022
* Mas i370 patch d (#383) * Refactor penciller memory In high-volume tests on large key-count clusters, so significant variation in the P0031 time has been seen: TimeBucket PatchA a.0ms_to_1ms 18554 b.1ms_to_2ms 51778 c.2ms_to_3ms 696 d.3ms_to_5ms 220 e.5ms_to_8ms 59 f.8ms_to_13ms 40 g.13ms_to_21ms 364 h.21ms_to_34ms 277 i.34ms_to_55ms 34 j.55ms_to_89ms 17 k.89ms_to_144ms 21 l.144ms_to_233ms 31 m.233ms_to_377ms 45 n.377ms_to_610ms 52 o.610ms_to_987ms 59 p.987ms_to_1597ms 55 q.1597ms_to_2684ms 54 r.2684ms_to_4281ms 29 s.4281ms_to_6965ms 7 t.6295ms_to_11246ms 1 It is unclear why this varies so much. The time to add to the cache appears to be minimal (but perhaps there is an issue with timing points in the code), whereas the time to add to the index is much more significant and variable. There is also variable time when the memory is rolled (although the actual activity here appears to be minimal. The refactoring here is two-fold: - tidy and simplify by keeping LoopState managed within handle_call, and add more helpful dialyzer specs; - change the update to the index to be a simple extension of a list, rather than any conversion. This alternative version of the pmem index in unit test is orders of magnitude faster to add - and is the same order of magnitude to check. Anticipation is that it may be more efficient in terms of memory changes. * Compress SST index Reduces the size of the leveled_sst index with two changes: 1 - Where there is a common prefix of tuple elements (e.g. Bucket) across the whole leveled_sst file - only the non-common part is indexed, and a function is used to compare. 2 - There is less "indexing" of the index i.e. only 1 in 16 keys are passed into the gb_trees part instead of 1 in 4 * Immediate hibernate Reasons for delay in hibernate were not clear. Straight after creation the process will not be in receipt of messages (must wait for the manifest to be updated), so better to hibernate now. This also means the log PC023 provides more accurate information. * Refactor BIC This patch avoids the following: - repeated replacement of the same element in the BIC (via get_kvrange), by checking presence via GET before sing SET - Stops re-reading of all elements to discover high modified date Also there appears to have been a bug where a missing HMD for the file is required to add to the cache. However, now the cache may be erased without erasing the HMD. This means that the cache can never be rebuilt * Use correct size in test results erts_debug:flat_size/1 returns size in words (i.e. 8 bytes on 64-bit CPU) not bytes * Don't change summary record As it is persisted as part of the file write, any change to the summary record cannot be rolled back * Clerk to prompt L0 write Simplifies the logic if the clerk request work for the penciller prompts L0 writes as well as Manifest changes. The advantage now is that if the penciller memory is full, and PUT load stops, the clerk should still be able to prompt persistence. the penciller can therefore make use of dead time this way * Add push on journal compact If there has been a backlog, followed by a quiet period - there may be a large ledger cache left unpushed. Journal compaction events are about once per hour, so the performance overhead of a false push should be minimal, with the advantage of clearing any backlog before load starts again. This is only relevant to riak users with very off/full batch type workloads. * Extend tests To more consistently trigger all overload scenarios * Fix range keys smaller than prefix Can't make end key an empty binary in this case, as it may be bigger than any keys within the range, but will appear to be smaller. Unit tests and ct tests added to expose the potential issue * Tidy-up - Remove penciller logs which are no longer called - Get pclerk to only wait MIN_TIMEOUT after doing work, in case there is a backlog - Remove update_levelzero_cache function as it is unique to handle_call of push_mem, and simple enough to be inline - Alight testutil slow offer with standard slow offer used * Tidy-up Remove pre-otp20 references. Reinstate the check that the starting pid is still active, this was added to tidy up shutdown. Resolve failure to run on otp20 due to `-if` sttaement * Tidy up Using null rather then {null, Key} is potentially clearer as it is not a concern what they Key is in this case, and removes a comparison step from the leveled_codec:endkey_passed/2 function. There were issues with coverage in eunit tests as the leveled_pclerk shut down. This prompted a general tidy of leveled_pclerk (remove passing of LoopState into internal functions, and add dialyzer specs. * Remove R16 relic * Further testing another issue The StartKey must always be less than or equal to the prefix when the first N characters are stripped, but this is not true of the EndKey (for the query) which does not have to be between the FirstKey and the LastKey. If the EndKey query does not match it must be greater than the Prefix (as otherwise it would not have been greater than the FirstKey - so set to null. * Fix unit test Unit test had a typo - and result interpretation had a misunderstanding. * Code and spec tidy Also look to the cover the situation when the FirstKey is the same as the Prefix with tests. This is, in theory, not an issue as it is the EndKey for each sublist which is indexed in leveled_tree. However, guard against it mapping to null here, just in case there are dangers lurking (note that tests will still pass without `M > N` guard in place. * Hibernate on BIC complete There are three situations when the BIC becomes complete: - In a file created as part of a merge the BIS is learned in the merge - After startup, files below L1 learn the block cache through reads that happen to read the block, eventually the while cache will be read, unless... - Either before/after the cache is complete, it can get whiped by a timeout after a get_sqn request (e.g. as prompted by a journal compaction) ... it will then be re-filled of the back of get/get-range requests. In all these situations we want to hibernate after the BIC is fill - to reflect the fact that the LoopState should now be relatively stable, so it is a good point to GC and rationalise location of data. Previously on the the first base was covered. Now all three are covered through the bic_complete message. * Test all index keys have same term This works functionally, but is not optimised (the term is replicated in the index) * Summaries with same index term If the summary index all have the same index term - only the object keys need to be indexes * Simplify case statements We either match the pattern of <<Prefix:N, Suffix>> or the answer should be null * OK for M == N If M = N for the first key, it will have a suffix of <<>>. This will match (as expected) a query Start Key of the sam size, and be smaller than any query Start Key that has the same prefix. If the query Start Key does not match the prefix - it will be null - as it must be smaller than the Prefix (as other wise the query Start Key would be bigger than the Last Key). The constraint of M > N was introduced before the *_prefix_filter functions were checking the prefix, to avoid issues. Now the prefix is being checked, then M == N is ok. * Simplify Correct the test to use a binary field in the range. To avoid further issue, only apply filter when everything is a binary() type. * Add test for head_only mode When leveled is used as a tictacaae key store (in parallel mode), the keys will be head_only entries. Double check they are handled as expected like object keys * Revert previous change - must support typed buckets Add assertion to confirm worthwhile optimisation * Add support for configurable cache multiple (#375) * Mas i370 patch e (#385) Improvement to monitoring for efficiency and improved readability of logs and stats. As part of this, where possible, tried to avoid updating loop state on READ messages in leveled processes (as was the case when tracking stats within each process). No performance benefits found with change, but improved stats has helped discover other potential gains.
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.
#370
Previous attempts to manage memory usage in large stores has focused on the Block Index Cache, which is about 2 bytes per key (for directly fetchable keys i.e. not index entries) in the leveled_sst files. These caches grow through use, but were never expunged - and perhaps files with mainly dead data would falsely retain these caches without performance benefits.
The balance is difficult though, between clearing these caches and the non-functional value in retaining them (especially wrt to increasing the speed of AAE-related queries).
Running a 36 hour test with relatively small values, to build up a key store.
Three things are noted:
process_info(P, memory)) is commonly about 8 times larger than the footprint of process state (as measured byerts_debug:flat_size(sys:get_state(P))) even after forced garbage collection on the PID. It is some times 25 times larger. This is the case even when binary memory owned by the process is empty (as measured byprocess_info(P, binary)). EDIT ... IGNORE THIS!erts_debug:flat_size/1is measured in words not bytes! One word equals 8 bytes on a 64-bit systemThis PR is a combination of three refactoring exercises intended to try and improve this situation:
[Addition|ExistingList]to extend lists).It is hoped that tidying up these elements will make it easier to investigate further memory overheads (and perhaps in the case of the leveled_sst index changes, directly improve memory footprint).