Relax the threshold for calling getDigestInExclusiveMode().

If the policy goal is to minimize disk seeks, this should be equally good.

Second attempt, with test update.

RELNOTES: None.
PiperOrigin-RevId: 202907857

1 files changed, 11 insertions, 7 deletions

diff --git a/src/main/java/com/google/devtools/build/lib/actions/cache/DigestUtils.java b/src/main/java/com/google/devtools/build/lib/actions/cache/DigestUtils.java
index 5e0f198273..37759f9ead 100644
--- a/src/main/java/com/google/devtools/build/lib/actions/cache/DigestUtils.java
+++ b/src/main/java/com/google/devtools/build/lib/actions/cache/DigestUtils.java
@@ -13,6 +13,7 @@
 // limitations under the License.
 package com.google.devtools.build.lib.actions.cache;
 
+import com.google.common.annotations.VisibleForTesting;
 import com.google.common.base.Preconditions;
 import com.google.common.cache.Cache;
 import com.google.common.cache.CacheBuilder;
@@ -59,6 +60,11 @@ public class DigestUtils {
   private static final Object DIGEST_LOCK = new Object();
   private static final AtomicBoolean MULTI_THREADED_DIGEST = new AtomicBoolean(false);
 
+  // Files of this size or less are assumed to be readable in one seek.
+  // (This is the default readahead window on Linux.)
+  @VisibleForTesting // the unittest is in a different package!
+  public static final int MULTI_THREADED_DIGEST_MAX_FILE_SIZE = 128 * 1024;
+
   // The time that a digest computation has to take at least in order to be considered a slow-read.
   private static final long SLOW_READ_MILLIS = 5000L;
 
@@ -238,13 +244,11 @@ public class DigestUtils {
 
     // All right, we have neither a fast nor a cached digest. Let's go through the costly process of
     // computing it from the file contents.
-    if (fileSize > 4096 && !MULTI_THREADED_DIGEST.get()) {
-      // We'll have to read file content in order to calculate the digest. In that case
-      // it would be beneficial to serialize those calculations since there is a high
-      // probability that MD5 will be requested for multiple output files simultaneously.
-      // Exception is made for small (<=4K) files since they will not likely to introduce
-      // significant delays (at worst they will result in two extra disk seeks by
-      // interrupting other reads).
+    if (fileSize > MULTI_THREADED_DIGEST_MAX_FILE_SIZE && !MULTI_THREADED_DIGEST.get()) {
+      // We'll have to read file content in order to calculate the digest.
+      // We avoid overlapping this process for multiple large files, as
+      // seeking back and forth between them will result in an overall loss of
+      // throughput.
       digest = getDigestInExclusiveMode(path);
     } else {
       digest = getDigestInternal(path);