IGNITE-23301 Fix scalability issues with coarse locks #4540
Conversation
ascherbakoff
changed the title
| // .jvmArgsAppend("-Djmh.executor=VIRTUAL") | ||
| // .addProfiler(JavaFlightRecorderProfiler.class, "configName=profile.jfc") |
There was a problem hiding this comment.
yes, let's keep it.
| * {@link ReadWriteLock#readLock()} operations at the cost of {@link ReadWriteLock#writeLock()} operations and memory consumption. It also | ||
| * supports reentrancy semantics like {@link ReentrantReadWriteLock}. | ||
| */ | ||
| public class StripedCompositeReadWriteLock implements ReadWriteLock { |
There was a problem hiding this comment.
You added a new concurrent primitive without any tests, could you pls add some?
| private long p0; | ||
|
|
||
| private long p1; | ||
|
|
||
| private long p2; | ||
|
|
||
| private long p3; | ||
|
|
||
| private long p4; | ||
|
|
||
| private long p5; | ||
|
|
||
| private long p6; | ||
|
|
||
| private long p7; |
There was a problem hiding this comment.
I've blindly copied it from AI2. Seems it's to prevent false sharing.
But it actually does nothing useful so I've removed it.
| assert false : "Unsupported coarse lock mode: " + lockMode; | ||
| } | ||
|
|
||
| return null; // Should not be here. |
There was a problem hiding this comment.
let's add assertion here
|
|
||
| UUID holderTx = slockOwners.keySet().iterator().next(); | ||
| CompletableFuture<Void> eventResult = fireEvent(LOCK_CONFLICT, new LockEventParameters(txId, holderTx)); | ||
| if (eventResult.isCompletedExceptionally()) { |
There was a problem hiding this comment.
pls add todo here: // TODO: https://issues.apache.org/jira/browse/IGNITE-21153
|
|
||
| break; | ||
| default: | ||
| assert false : "Unsupported coarse unlock mode: " + lock.lockMode(); |
There was a problem hiding this comment.
seems SIX lock mode is also possible here
There was a problem hiding this comment.
No, in the current implementation SIX lock is acquired by holding two locks: IX + S.
There was a problem hiding this comment.
Agree, I should have checked more carefully.
modules/core/src/main/java/org/apache/ignite/internal/util/StripedCompositeReadWriteLock.java
Outdated
Show resolved
Hide resolved
...table/src/main/java/org/apache/ignite/internal/table/distributed/raft/PartitionListener.java
Show resolved
Hide resolved
...n/java/org/apache/ignite/internal/table/distributed/replicator/PartitionReplicaListener.java
Show resolved
Hide resolved
ascherbakoff
force-pushed
the
ignite-23301
branch
from
b6d4c0f
to
755ef13
Compare
modules/transactions/src/main/java/org/apache/ignite/internal/tx/impl/HeapLockManager.java
Show resolved
Hide resolved
modules/core/src/main/java/org/apache/ignite/internal/util/IgniteStripedReadWriteLock.java
Show resolved
Hide resolved
modules/transactions/src/main/java/org/apache/ignite/internal/tx/impl/HeapLockManager.java
Outdated
Show resolved
Hide resolved
| null, | ||
| indexIdsAtRwTxBeginTs(txId) | ||
| ); | ||
| if (!IgniteSystemProperties.getBoolean(IgniteSystemProperties.IGNITE_SKIP_STORAGE_UPDATE_IN_BENCHMARK)) { |
There was a problem hiding this comment.
let's cache this to avoid lookup in System.getProperty()
| // Validate reordering with IX locks. | ||
| for (Lock lock : ixlockOwners.values()) { | ||
| // Allow only high priority transactions to wait. | ||
| if (TxIdPriorityComparator.INSTANCE.compare(lock.txId(), txId) < 0) { |
There was a problem hiding this comment.
TxIdComparator from deadlockPreventionPolicy should be used here, otherwise we will have lock conflicts if we change the tx id ordering in the deadlock prevention policy.
https://issues.apache.org/jira/browse/IGNITE-23301
Implemented scalable coarse (table) locks.
It only supports IX, S, and SIX lock modes. SIX mode is not avaiable directly, only by upgrading IX->S or S->IX
Short locks are not supported
IX locks are optimized for concurrency.
Implementation is deadlock-free.
Table locks now are partition locks.
IS coarse locks removed because they are not needed.
Benchmark results (Xeon Silver 4314 CPU):
old ver (rev 33c1649)
1 thread
Benchmark (batch) (fsync) (partitionCount) Mode Cnt Score Error Units
UpsertKvBenchmark.upsert 1 false 8 thrpt 20 68650.803 ± 6967.343 ops/s
16 threads
Benchmark (batch) (fsync) (partitionCount) Mode Cnt Score Error Units
UpsertKvBenchmark.upsert 1 false 8 thrpt 20 202157.172 ± 10789.938 ops/s
32 threads
Benchmark (batch) (fsync) (partitionCount) Mode Cnt Score Error Units
UpsertKvBenchmark.upsert 1 false 8 thrpt 20 167636.692 ± 12318.723 ops/s
new ver
1 thread
Benchmark (batch) (fsync) (partitionCount) Mode Cnt Score Error Units
UpsertKvBenchmark.upsert 1 false 8 thrpt 20 74008.938 ± 5447.381 ops/s
16 threads
Benchmark (batch) (fsync) (partitionCount) Mode Cnt Score Error Units
UpsertKvBenchmark.upsert 1 false 8 thrpt 20 290628.268 ± 11657.284 ops/s
32 threads
Benchmark (batch) (fsync) (partitionCount) Mode Cnt Score Error Units
UpsertKvBenchmark.upsert 1 false 8 thrpt 20 320520.489 ± 16084.668 ops/s