Issue 3645: Allow direct reads in RocksDB to be configurable. (praveg…

…a#3647) Allow RocksDB direct reads to be configurable. Signed-off-by: Raúl Gracia <[email protected]>
vijikarthi · Apr 17, 2019 · a0ed24e · a0ed24e
1 parent b30f2bd
commit a0ed24e
Show file tree

Hide file tree

Showing 3 changed files with 19 additions and 1 deletion.
diff --git a/config/config.properties b/config/config.properties
@@ -405,6 +405,12 @@ extendeds3.url=localhost:9020
 # will decrease, so the index size will also reduce linearly (but increasing read amplification).
 #rocksdb.cacheBlockSizeKB=32
 
+# According to RocksDB documentation, enabling direct reads may be beneficial for performance due to: i) it avoids extra
+# copies of data on OS page cache, ii) it exploits better knowledge of the behavior of data to apply policies (e.g.,
+# replacement). However, as not all OS/environments support direct IO, we keep it disabled by default for safety. For
+# more information, please check: https://github.com/facebook/rocksdb/wiki/Direct-IO.
+#rocksdb.directReads=false
+
 ##endregion
 
 ##region DurableLog Settings

diff --git a/...storage/impl/src/main/java/io/pravega/segmentstore/storage/impl/rocksdb/RocksDBCache.java b/...storage/impl/src/main/java/io/pravega/segmentstore/storage/impl/rocksdb/RocksDBCache.java
@@ -69,6 +69,7 @@ class RocksDBCache implements Cache {
     private final int writeBufferSizeMB;
     private final int readCacheSizeMB;
     private final int cacheBlockSizeKB;
+    private final boolean directReads;
 
     //endregion
 
@@ -95,6 +96,7 @@ class RocksDBCache implements Cache {
         this.writeBufferSizeMB = config.getWriteBufferSizeMB() / MAX_WRITE_BUFFER_NUMBER;
         this.readCacheSizeMB = config.getReadCacheSizeMB();
         this.cacheBlockSizeKB = config.getCacheBlockSizeKB();
+        this.directReads = config.isDirectReads();
         try {
             this.databaseOptions = createDatabaseOptions();
             this.writeOptions = createWriteOptions();
@@ -254,7 +256,7 @@ private Options createDatabaseOptions() {
                 .setMinWriteBufferNumberToMerge(MIN_WRITE_BUFFER_NUMBER_TO_MERGE)
                 .setTableFormatConfig(tableFormatConfig)
                 .setOptimizeFiltersForHits(true)
-                .setUseDirectReads(true);
+                .setUseDirectReads(this.directReads);
     }
 
     private void clear(boolean recreateDirectory) {

diff --git a/...torage/impl/src/main/java/io/pravega/segmentstore/storage/impl/rocksdb/RocksDBConfig.java b/...torage/impl/src/main/java/io/pravega/segmentstore/storage/impl/rocksdb/RocksDBConfig.java
@@ -25,6 +25,7 @@ public class RocksDBConfig {
     public static final Property<Integer> WRITE_BUFFER_SIZE_MB = Property.named("writeBufferSizeMB", 64);
     public static final Property<Integer> READ_CACHE_SIZE_MB = Property.named("readCacheSizeMB", 8);
     public static final Property<Integer> CACHE_BLOCK_SIZE_KB = Property.named("cacheBlockSizeKB", 32);
+    public static final Property<Boolean> DIRECT_READS = Property.named("directReads", false);
     private static final String COMPONENT_CODE = "rocksdb";
 
     //endregion
@@ -60,6 +61,14 @@ public class RocksDBConfig {
     @Getter
     private final int cacheBlockSizeKB;
 
+    /**
+     * Enabling direct reads may be beneficial for performance due to: i) it avoids extra copies of data on OS page
+     * cache, ii) it exploits better knowledge of the behavior of data to apply policies (e.g., replacement). However,
+     * as not all OS/environments support direct IO, we keep it disabled by default for safety.
+     */
+    @Getter
+    private final boolean directReads;
+
     //endregion
 
     //region Constructor
@@ -74,6 +83,7 @@ private RocksDBConfig(TypedProperties properties) throws ConfigurationException
         this.writeBufferSizeMB = properties.getInt(WRITE_BUFFER_SIZE_MB);
         this.readCacheSizeMB = properties.getInt(READ_CACHE_SIZE_MB);
         this.cacheBlockSizeKB = properties.getInt(CACHE_BLOCK_SIZE_KB);
+        this.directReads = properties.getBoolean(DIRECT_READS);
     }
 
     /**