apple/foundationdb
1
0
Fork 0
mirror of https://github.com/apple/foundationdb.git synced 2025-06-02 19:25:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #1527 from alexmiller-apple/tstlog-6.1

Browse Source 
Spill-by-reference knob + TLog6.0 Spilled Peek deprioritization
This commit is contained in:
Evan Tschannen 2019-05-03 17:19:45 -07:00 committed by GitHub
commit 93eb2a9395
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 29 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
fdbserver/Knobs.cpp
View File

@ -72,6 +72,7 @@ ServerKnobs::ServerKnobs(bool randomize, ClientKnobs* clientKnobs) {
init( DISK_QUEUE_ADAPTER_MAX_SWITCH_TIME, 5.0 );
init( TLOG_SPILL_REFERENCE_MAX_PEEK_MEMORY_BYTES, 2e9 ); if ( randomize && BUGGIFY ) TLOG_SPILL_REFERENCE_MAX_PEEK_MEMORY_BYTES = 2e6;
init( TLOG_SPILL_REFERENCE_MAX_BATCHES_PER_PEEK, 100 ); if ( randomize && BUGGIFY ) TLOG_SPILL_REFERENCE_MAX_BATCHES_PER_PEEK = 1;
init( TLOG_SPILL_REFERENCE_MAX_BYTES_PER_BATCH, 16<<10 ); if ( randomize && BUGGIFY ) TLOG_SPILL_REFERENCE_MAX_BYTES_PER_BATCH = 500;
init( DISK_QUEUE_FILE_EXTENSION_BYTES, 10<<20 ); // BUGGIFYd per file within the DiskQueue
init( DISK_QUEUE_FILE_SHRINK_BYTES, 100<<20 ); // BUGGIFYd per file within the DiskQueue
init( TLOG_DEGRADED_DELAY_COUNT, 5 );

1
fdbserver/Knobs.h
View File

@ -76,6 +76,7 @@ public:
double DISK_QUEUE_ADAPTER_MAX_SWITCH_TIME;
int64_t TLOG_SPILL_REFERENCE_MAX_PEEK_MEMORY_BYTES;
int64_t TLOG_SPILL_REFERENCE_MAX_BATCHES_PER_PEEK;
int64_t TLOG_SPILL_REFERENCE_MAX_BYTES_PER_BATCH;
int64_t DISK_QUEUE_FILE_EXTENSION_BYTES; // When we grow the disk queue, by how many bytes should it grow?
int64_t DISK_QUEUE_FILE_SHRINK_BYTES; // When we shrink the disk queue, by how many bytes should it shrink?
int TLOG_DEGRADED_DELAY_COUNT;

9
fdbserver/OldTLogServer_6_0.actor.cpp
View File

@ -996,6 +996,15 @@ ACTOR Future<Void> tLogPeekMessages( TLogData* self, TLogPeekRequest req, Refere
wait( delay(0.0, TaskLowPriority) );
}
if( req.begin <= logData->persistentDataDurableVersion && req.tag != txsTag) {
// Reading spilled data will almost always imply that the storage server is >5s behind the rest
// of the cluster. We shouldn't prioritize spending CPU on helping this server catch up
// slightly faster over keeping the rest of the cluster operating normally.
// txsTag is only ever peeked on recovery, and we would still wish to prioritize requests
// that impact recovery duration.
wait(delay(0, TaskTLogSpilledPeekReply));
}
Version poppedVer = poppedVersion(logData, req.tag);
if(poppedVer > req.begin) {
TLogPeekReply rep;

23
fdbserver/TLogServer.actor.cpp
View File

@ -795,6 +795,15 @@ ACTOR Future<Void> updatePersistentData( TLogData* self, Reference<LogData> logD
lastVersion = std::max(currentVersion, lastVersion);
firstLocation = std::min(begin, firstLocation);
if ((wr.getLength() + sizeof(SpilledData) > SERVER_KNOBS->TLOG_SPILL_REFERENCE_MAX_BYTES_PER_BATCH) ) {
*(uint32_t*)wr.getData() = refSpilledTagCount;
self->persistentData->set( KeyValueRef( persistTagMessageRefsKey( logData->logId, tagData->tag, lastVersion ), wr.toValue() ) );
tagData->poppedLocation = std::min(tagData->poppedLocation, firstLocation);
refSpilledTagCount = 0;
wr = BinaryWriter( AssumeVersion(logData->protocolVersion) );
wr << uint32_t(0);
}
Future<Void> f = yield(TaskUpdateStorage);
if(!f.isReady()) {
wait(f);
@ -1277,6 +1286,15 @@ ACTOR Future<Void> tLogPeekMessages( TLogData* self, TLogPeekRequest req, Refere
wait( delay(0.0, TaskLowPriority) );
}
if( req.begin <= logData->persistentDataDurableVersion && req.tag != txsTag) {
// Reading spilled data will almost always imply that the storage server is >5s behind the rest
// of the cluster. We shouldn't prioritize spending CPU on helping this server catch up
// slightly faster over keeping the rest of the cluster operating normally.
// txsTag is only ever peeked on recovery, and we would still wish to prioritize requests
// that impact recovery duration.
wait(delay(0, TaskTLogSpilledPeekReply));
}
Version poppedVer = poppedVersion(logData, req.tag);
if(poppedVer > req.begin) {
TLogPeekReply rep;
@ -1338,11 +1356,6 @@ ACTOR Future<Void> tLogPeekMessages( TLogData* self, TLogPeekRequest req, Refere
else
messages.serializeBytes( messages2.toValue() );
} else {
// Calculating checksums of read pages is potentially expensive, and storage servers with
// spilled data are likely behind and not contributing usefully to the cluster anyway.
// Thus, we penalize their priority slightly to make sure that commits have a higher priority
// than catching up old storage servers.
wait(delay(0, TaskTLogSpilledPeekReply));
// FIXME: Limit to approximately DESIRED_TOTATL_BYTES somehow.
Standalone<VectorRef<KeyValueRef>> kvrefs = wait(
self->persistentData->readRange(KeyRangeRef(