Merge pull request #761 from IntersectMBO/dcoutts/wp8-bench-pipelined-3

Improve parallel speedup on pipelined benchmark

Author: jorisdral

README.md

@@ -356,6 +356,12 @@ The *disk cache policy* determines if lookup operations use the OS page
 cache. Caching may improve the performance of lookups and updates if
 database access follows certain patterns.
 
+`confMergeBatchSize`  
+The merge batch size balances the maximum latency of individual update
+operations, versus the latency of a sequence of update operations.
+Bigger batches improves overall performance but some updates will take a
+lot longer than others. The default is to use a large batch size.
+
 ##### Fine-tuning: Merge Policy, Size Ratio, and Write Buffer Size <span id="fine_tuning_data_layout" class="anchor"></span>
 
 The configuration parameters `confMergePolicy`, `confSizeRatio`, and
@@ -647,6 +653,33 @@ locality if it is likely to access entries that have nearby keys.
   does not have good spatial or temporal locality. For instance, if the
   access pattern is uniformly random.
 
+##### Fine-tuning: Merge Batch Size <span id="fine_tuning_merge_batch_size" class="anchor"></span>
+
+The *merge batch size* is a micro-tuning parameter, and in most cases
+you do need to think about it and can leave it at its default.
+
+When using the `Incremental` merge schedule, merging is done in batches.
+This is a trade-off: larger batches tends to mean better overall
+performance but the downside is that while most updates (inserts,
+deletes, upserts) are fast, some are slower (when a batch of merging
+work has to be done).
+
+If you care most about the maximum latency of updates, then use a small
+batch size. If you don't care about latency of individual operations,
+just the latency of the overall sequence of operations then use a large
+batch size. The default is to use a large batch size, the same size as
+the write buffer itself. The minimum batch size is 1. The maximum batch
+size is the size of the write buffer `confWriteBufferAlloc`.
+
+Note that the actual batch size is the minimum of this configuration
+parameter and the size of the batch of operations performed (e.g.
+`inserts`). So if you consistently use large batches, you can use a
+batch size of 1 and the merge batch size will always be determined by
+the operation batch size.
+
+A further reason why it may be preferable to use minimal batch sizes is
+to get good parallel work balance, when using parallelism.
+
 ### References
 
 The implementation of LSM-trees in this package draws inspiration from:

bench/macro/lsm-tree-bench-wp8.hs

@@ -183,13 +183,23 @@ mkTableConfigSetup GlobalOpts{diskCachePolicy} SetupOpts{bloomFilterAlloc} conf
     , LSM.confBloomFilterAlloc = bloomFilterAlloc
     }
 
-mkTableConfigRun :: GlobalOpts -> LSM.TableConfig -> LSM.TableConfig
-mkTableConfigRun GlobalOpts{diskCachePolicy} conf = conf {
-      LSM.confDiskCachePolicy = diskCachePolicy
+mkTableConfigRun :: GlobalOpts -> RunOpts -> LSM.TableConfig -> LSM.TableConfig
+mkTableConfigRun GlobalOpts{diskCachePolicy} RunOpts {pipelined} conf =
+    conf {
+      LSM.confDiskCachePolicy = diskCachePolicy,
+      LSM.confMergeBatchSize  = if pipelined
+                                  then LSM.MergeBatchSize 1
+                                  else LSM.confMergeBatchSize conf
     }
 
-mkOverrideDiskCachePolicy :: GlobalOpts -> LSM.OverrideDiskCachePolicy
-mkOverrideDiskCachePolicy GlobalOpts{diskCachePolicy} = LSM.OverrideDiskCachePolicy diskCachePolicy
+mkTableConfigOverride :: GlobalOpts -> RunOpts -> LSM.TableConfigOverride
+mkTableConfigOverride GlobalOpts{diskCachePolicy} RunOpts {pipelined} =
+    LSM.noTableConfigOverride {
+      LSM.overrideDiskCachePolicy = Just diskCachePolicy,
+      LSM.overrideMergeBatchSize  = if pipelined
+                                      then Just (LSM.MergeBatchSize 1)
+                                      else Nothing
+    }
 
 mkTracer :: GlobalOpts -> Tracer IO LSM.LSMTreeTrace
 mkTracer gopts
@@ -585,8 +595,10 @@ doRun gopts opts = do
         -- reference version starts with empty (as it's not practical or
         -- necessary for testing to load the whole snapshot).
         tbl <- if check opts
-                then LSM.newTableWith @IO @K @V @B (mkTableConfigRun gopts benchTableConfig) session
-                else LSM.openTableFromSnapshotWith @IO @K @V @B (mkOverrideDiskCachePolicy gopts) session name label
+                then let conf = mkTableConfigRun gopts opts benchTableConfig
+                      in LSM.newTableWith @IO @K @V @B conf session
+                else let conf = mkTableConfigOverride gopts opts
+                      in LSM.openTableFromSnapshotWith @IO @K @V @B conf session name label
 
         -- In checking mode, compare each output against a pure reference.
         checkvar <- newIORef $ pureReference

lsm-tree.cabal

@@ -183,6 +183,12 @@ description:
       The /disk cache policy/ determines if lookup operations use the OS page cache.
       Caching may improve the performance of lookups and updates if database access follows certain patterns.
 
+  [@confMergeBatchSize@]
+      The merge batch size balances the maximum latency of individual update
+      operations, versus the latency of a sequence of update operations. Bigger
+      batches improves overall performance but some updates will take a lot
+      longer than others. The default is to use a large batch size.
+
   ==== Fine-tuning: Merge Policy, Size Ratio, and Write Buffer Size #fine_tuning_data_layout#
 
   The configuration parameters @confMergePolicy@, @confSizeRatio@, and @confWriteBufferAlloc@ affect how the table organises its data.
@@ -429,6 +435,31 @@ description:
   *   Use the @DiskCacheNone@ policy if the database's access pattern has does not have good spatial or temporal locality.
       For instance, if the access pattern is uniformly random.
 
+  ==== Fine-tuning: Merge Batch Size #fine_tuning_merge_batch_size#
+
+  The /merge batch size/ is a micro-tuning parameter, and in most cases you do
+  need to think about it and can leave it at its default.
+
+  When using the 'Incremental' merge schedule, merging is done in batches. This
+  is a trade-off: larger batches tends to mean better overall performance but the
+  downside is that while most updates (inserts, deletes, upserts) are fast, some
+  are slower (when a batch of merging work has to be done).
+
+  If you care most about the maximum latency of updates, then use a small batch
+  size. If you don't care about latency of individual operations, just the
+  latency of the overall sequence of operations then use a large batch size. The
+  default is to use a large batch size, the same size as the write buffer itself.
+  The minimum batch size is 1. The maximum batch size is the size of the write
+  buffer 'confWriteBufferAlloc'.
+
+  Note that the actual batch size is the minimum of this configuration
+  parameter and the size of the batch of operations performed (e.g. 'inserts').
+  So if you consistently use large batches, you can use a batch size of 1 and
+  the merge batch size will always be determined by the operation batch size.
+
+  A further reason why it may be preferable to use minimal batch sizes is to get
+  good parallel work balance, when using parallelism.
+
   == References
 
   The implementation of LSM-trees in this package draws inspiration from:

src-extras/Database/LSMTree/Extras/NoThunks.hs

@@ -659,6 +659,9 @@ deriving anyclass instance NoThunks DiskCachePolicy
 deriving stock instance Generic MergeSchedule
 deriving anyclass instance NoThunks MergeSchedule
 
+deriving stock instance Generic MergeBatchSize
+deriving anyclass instance NoThunks MergeBatchSize
+
 {-------------------------------------------------------------------------------
   RWVar
 -------------------------------------------------------------------------------}

src/Database/LSMTree.hs

@@ -116,7 +116,8 @@ module Database.LSMTree (
     confBloomFilterAlloc,
     confFencePointerIndex,
     confDiskCachePolicy,
-    confMergeSchedule
+    confMergeSchedule,
+    confMergeBatchSize
   ),
   defaultTableConfig,
   MergePolicy (LazyLevelling),
@@ -126,9 +127,11 @@ module Database.LSMTree (
   BloomFilterAlloc (AllocFixed, AllocRequestFPR),
   FencePointerIndexType (OrdinaryIndex, CompactIndex),
   DiskCachePolicy (..),
+  MergeBatchSize (..),
 
   -- ** Table Configuration Overrides #table_configuration_overrides#
-  OverrideDiskCachePolicy (..),
+  TableConfigOverride (..),
+  noTableConfigOverride,
 
   -- * Ranges #ranges#
   Range (..),
@@ -221,11 +224,12 @@ import qualified Database.LSMTree.Internal.BlobRef as Internal
 import           Database.LSMTree.Internal.Config
                      (BloomFilterAlloc (AllocFixed, AllocRequestFPR),
                      DiskCachePolicy (..), FencePointerIndexType (..),
-                     LevelNo (..), MergePolicy (..), MergeSchedule (..),
-                     SizeRatio (..), TableConfig (..), WriteBufferAlloc (..),
-                     defaultTableConfig, serialiseKeyMinimalSize)
+                     LevelNo (..), MergeBatchSize (..), MergePolicy (..),
+                     MergeSchedule (..), SizeRatio (..), TableConfig (..),
+                     WriteBufferAlloc (..), defaultTableConfig,
+                     serialiseKeyMinimalSize)
 import           Database.LSMTree.Internal.Config.Override
-                     (OverrideDiskCachePolicy (..))
+                     (TableConfigOverride (..), noTableConfigOverride)
 import           Database.LSMTree.Internal.Entry (NumEntries (..))
 import qualified Database.LSMTree.Internal.Entry as Entry
 import           Database.LSMTree.Internal.Merge (LevelMergeType (..))
@@ -2600,7 +2604,7 @@ Variant of 'withTableFromSnapshot' that accepts [table configuration overrides](
   withTableFromSnapshotWith ::
     forall k v b a.
     (ResolveValue v) =>
-    OverrideDiskCachePolicy ->
+    TableConfigOverride ->
     Session IO ->
     SnapshotName ->
     SnapshotLabel ->
@@ -2611,7 +2615,7 @@ withTableFromSnapshotWith ::
   forall m k v b a.
   (IOLike m) =>
   (ResolveValue v) =>
-  OverrideDiskCachePolicy ->
+  TableConfigOverride ->
   Session m ->
   SnapshotName ->
   SnapshotLabel ->
@@ -2675,7 +2679,7 @@ openTableFromSnapshot ::
   SnapshotLabel ->
   m (Table m k v b)
 openTableFromSnapshot session snapName snapLabel =
-  openTableFromSnapshotWith NoOverrideDiskCachePolicy session snapName snapLabel
+  openTableFromSnapshotWith noTableConfigOverride session snapName snapLabel
 
 {- |
 Variant of 'openTableFromSnapshot' that accepts [table configuration overrides](#g:table_configuration_overrides).
@@ -2684,7 +2688,7 @@ Variant of 'openTableFromSnapshot' that accepts [table configuration overrides](
   openTableFromSnapshotWith ::
     forall k v b.
     (ResolveValue v) =>
-    OverrideDiskCachePolicy ->
+    TableConfigOverride ->
     Session IO ->
     SnapshotName ->
     SnapshotLabel ->
@@ -2694,7 +2698,7 @@ openTableFromSnapshotWith ::
   forall m k v b.
   (IOLike m) =>
   (ResolveValue v) =>
-  OverrideDiskCachePolicy ->
+  TableConfigOverride ->
   Session m ->
   SnapshotName ->
   SnapshotLabel ->

src/Database/LSMTree/Internal/Config.hs

@@ -26,12 +26,16 @@ module Database.LSMTree.Internal.Config (
   , diskCachePolicyForLevel
     -- * Merge schedule
   , MergeSchedule (..)
+    -- * Merge batch size
+  , MergeBatchSize (..)
+  , creditThresholdForLevel
   ) where
 
 import           Control.DeepSeq (NFData (..))
 import           Database.LSMTree.Internal.Index (IndexType)
 import qualified Database.LSMTree.Internal.Index as Index
                      (IndexType (Compact, Ordinary))
+import qualified Database.LSMTree.Internal.MergingRun as MR
 import qualified Database.LSMTree.Internal.RawBytes as RB
 import           Database.LSMTree.Internal.Run (RunDataCaching (..))
 import           Database.LSMTree.Internal.RunAcc (RunBloomFilterAlloc (..))
@@ -90,6 +94,12 @@ For a detailed discussion of fine-tuning the table configuration, see [Fine-tuni
 [@confDiskCachePolicy :: t'DiskCachePolicy'@]
     The /disk cache policy/ supports caching lookup operations using the OS page cache.
     Caching may improve the performance of lookups and updates if database access follows certain patterns.
+
+[@confMergeBatchSize :: t'MergeBatchSize'@]
+    The merge batch size balances the maximum latency of individual update
+    operations, versus the latency of a sequence of update operations. Bigger
+    batches improves overall performance but some updates will take a lot
+    longer than others. The default is to use a large batch size.
 -}
 data TableConfig = TableConfig {
     confMergePolicy       :: !MergePolicy
@@ -99,12 +109,14 @@ data TableConfig = TableConfig {
   , confBloomFilterAlloc  :: !BloomFilterAlloc
   , confFencePointerIndex :: !FencePointerIndexType
   , confDiskCachePolicy   :: !DiskCachePolicy
+  , confMergeBatchSize    :: !MergeBatchSize
   }
   deriving stock (Show, Eq)
 
 instance NFData TableConfig where
-  rnf (TableConfig a b c d e f g) =
-      rnf a `seq` rnf b `seq` rnf c `seq` rnf d `seq` rnf e `seq` rnf f `seq` rnf g
+  rnf (TableConfig a b c d e f g h) =
+      rnf a `seq` rnf b `seq` rnf c `seq` rnf d `seq`
+      rnf e `seq` rnf f `seq` rnf g `seq` rnf h
 
 -- | The 'defaultTableConfig' defines reasonable defaults for all 'TableConfig' parameters.
 --
@@ -122,6 +134,8 @@ instance NFData TableConfig where
 -- OrdinaryIndex
 -- >>> confDiskCachePolicy defaultTableConfig
 -- DiskCacheAll
+-- >>> confMergeBatchSize defaultTableConfig
+-- MergeBatchSize 20000
 --
 defaultTableConfig :: TableConfig
 defaultTableConfig =
@@ -133,6 +147,7 @@ defaultTableConfig =
       , confBloomFilterAlloc  = AllocRequestFPR 1.0e-3
       , confFencePointerIndex = OrdinaryIndex
       , confDiskCachePolicy   = DiskCacheAll
+      , confMergeBatchSize    = MergeBatchSize 20_000 -- same as write buffer
       }
 
 data RunLevelNo = RegularLevel LevelNo | UnionLevel
@@ -238,6 +253,8 @@ data MergeSchedule =
     The 'Incremental' merge schedule spreads out the merging work over time.
     This is less efficient than the 'OneShot' merge schedule, but has a consistent workload.
     Using the 'Incremental' merge schedule, the worst-case disk I\/O complexity of the update operations is /logarithmic/ in the size of the table.
+    This 'Incremental' merge schedule still uses batching to improve performance.
+    The batch size can be controlled using the 'MergeBatchSize'.
     -}
   | Incremental
   deriving stock (Eq, Show)
@@ -385,3 +402,50 @@ diskCachePolicyForLevel policy levelNo =
         RegularLevel l | l <= LevelNo n -> CacheRunData
                        | otherwise      -> NoCacheRunData
         UnionLevel                      -> NoCacheRunData
+
+{-------------------------------------------------------------------------------
+  Merge batch size
+-------------------------------------------------------------------------------}
+
+{- |
+The /merge batch size/ is a micro-tuning parameter, and in most cases you do
+need to think about it and can leave it at its default.
+
+When using the 'Incremental' merge schedule, merging is done in batches. This
+is a trade-off: larger batches tends to mean better overall performance but the
+downside is that while most updates (inserts, deletes, upserts) are fast, some
+are slower (when a batch of merging work has to be done).
+
+If you care most about the maximum latency of updates, then use a small batch
+size. If you don't care about latency of individual operations, just the
+latency of the overall sequence of operations then use a large batch size. The
+default is to use a large batch size, the same size as the write buffer itself.
+The minimum batch size is 1. The maximum batch size is the size of the write
+buffer 'confWriteBufferAlloc'.
+
+Note that the actual batch size is the minimum of this configuration
+parameter and the size of the batch of operations performed (e.g. 'inserts').
+So if you consistently use large batches, you can use a batch size of 1 and
+the merge batch size will always be determined by the operation batch size.
+
+A further reason why it may be preferable to use minimal batch sizes is to get
+good parallel work balance, when using parallelism.
+-}
+newtype MergeBatchSize = MergeBatchSize Int
+  deriving stock (Show, Eq, Ord)
+  deriving newtype (NFData)
+
+-- TODO: the thresholds for doing merge work should be different for each level,
+-- and ideally all-pairs co-prime.
+creditThresholdForLevel :: TableConfig -> LevelNo -> MR.CreditThreshold
+creditThresholdForLevel TableConfig {
+                           confMergeBatchSize   = MergeBatchSize mergeBatchSz,
+                           confWriteBufferAlloc = AllocNumEntries writeBufferSz
+                         }
+                        (LevelNo _i) =
+    MR.CreditThreshold
+  . MR.UnspentCredits
+  . MR.MergeCredits
+  . max 1
+  . min writeBufferSz
+  $ mergeBatchSz