[JGIT PATCH 01/23] Improve hit performance on the UnpackedObjectCache
- From
Shawn O. Pearce <spearce@spearce.org>
- Date
- Dec 25, 2008, 02:10 UTC
- Message-ID
- <1230171079-17156-2-git-send-email-spearce@spearce.org>
- In-Reply-To
- <1230171079-17156-1-git-send-email-spearce@spearce.org>
If the JVM cleared one of our SoftReferences we "leaked" the space in the cache that the entry occupied. Over time this meant we lost room in the cache and didn't have a way to recover that when we replaced the evicted entry.
The hash function was also too complex for the hit ratio we were getting. The old function on one of my linux-2.6 clones was giving us <7% hit ratio; this new function is a little simpler to compute and is getting ~11%. Increasing the size of the hash table helps matters considerably.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org> --- .../org/spearce/jgit/lib/UnpackedObjectCache.java | 21 +++++++++---------- 1 files changed, 10 insertions(+), 11 deletions(-)
diff --git a/org.spearce.jgit/src/org/spearce/jgit/lib/UnpackedObjectCache.java b/org.spearce.jgit/src/org/spearce/jgit/lib/UnpackedObjectCache.java index ee6a680..677b3a7 100644 --- a/org.spearce.jgit/src/org/spearce/jgit/lib/UnpackedObjectCache.java +++ b/org.spearce.jgit/src/org/spearce/jgit/lib/UnpackedObjectCache.java @@ -40,17 +40,14 @@ import java.lang.ref.SoftReference; class UnpackedObjectCache { - private static final int CACHE_SZ = 256; + private static final int CACHE_SZ = 1024; private static final int MB = 1024 * 1024; private static final SoftReference<Entry> DEAD; - private static int hash(final WindowedFile pack, final long position) { - int h = pack.hash + (int) position; - h += h >>> 16; - h += h >>> 8; - return h & (CACHE_SZ - 1); + private static int hash(final long position) { + return (((int) position) << 22) >>> 22; } private static int maxByteCount; @@ -80,7 +77,7 @@ static synchronized void reconfigure(final int dbLimit) { } static synchronized Entry get(final WindowedFile pack, final long position) { - final Slot e = cache[hash(pack, position)]; + final Slot e = cache[hash(position)]; if (e.provider == pack && e.position == position) { final Entry buf = e.data.get(); if (buf != null) { @@ -96,7 +93,7 @@ static synchronized void store(final WindowedFile pack, if (data.length > maxByteCount) return; // Too large to cache. - final Slot e = cache[hash(pack, position)]; + final Slot e = cache[hash(position)]; clearEntry(e); openByteCount += data.length; @@ -104,6 +101,7 @@ static synchronized void store(final WindowedFile pack, e.provider = pack; e.position = position; + e.sz = data.length; e.data = new SoftReference<Entry>(new Entry(data, objectType)); moveToHead(e); } @@ -155,11 +153,10 @@ private static void unlink(final Slot e) { } private static void clearEntry(final Slot e) { - final Entry old = e.data.get(); - if (old != null) - openByteCount -= old.data.length; + openByteCount -= e.sz; e.provider = null; e.data = DEAD; + e.sz = 0; } private UnpackedObjectCache() { @@ -186,6 +183,8 @@ Entry(final byte[] aData, final int aType) { long position; + int sz; + SoftReference<Entry> data = DEAD; } }
-- 1.6.1.rc4.301.g5497a