WebCore:
Reviewed by Dave Hyatt.
Tweaked the cache eviction model to better balance between live and
dead resources.
For the sake of avoiding evictions during the PLT, the old model
required the sum of dead and live resources to grow to twice the cache
capacity before evicting, and would then evict dead or live down to 0
if necessary. This was a too-high high water mark, which would nullify
much of the value of eviction, and a too-low low water mark, which
would nullify much of the value of the LRU-SP strategy.
This patch changes the model in 3 ways.
1. The new model for dead resources is a flexible window with a fixed
minimum and maximum. The dead resource window is big when live resource
pressure is small, and vice versa. This has the immediate advantage of
cutting the high water mark by up to 50%. It also enables the following
tunable optimizations in future patches:
a. A dead resource limit of 0 for clients who want that. (Just set
the fixed maximum to 0.)
b. A much higher low water mark. (Just set the fixed minimum to,
say, 25% of the cache's capacity.)
c. A much lower high water mark for users who browse simple pages
in one tab. (Just set the fixed maximum to, say, 50% of the cache's
capacity.)
I plan to make the changes that actually take advantage of these
tunable optimizations in another check-in.
The new model won't hurt the PLT because it will notice the PLT's low
live resource size, and up the dead resource capacity in response. For
the same reason, the new model should establish a good balance in
real-world use.
2. Live resource eviction is now based on size(), not encodedSize().
So, a page with lots of large, encoded images will start evicting
resources, if necessary, even before all the images paint. This allows
you to more accurately stipulate an exact high water mark.
3. When pruning, prune to a small percentage below capacity, to avoid
just having to prune again immediately.
Layout tests pass. PLT shows no regression.
* history/PageCache.cpp:
(WebCore::PageCache::releaseAutoreleasedPagesNow): Updated for rename.
* loader/Cache.cpp: Implemented the algorithm explained above.
* loader/Cache.h: Removed explicit tracking of decoded data size, since
it was unused.
* loader/CachedResource.cpp: ditto on tracking of decoded data size
WebKit:
Reviewed by Dave Hyatt.
WebKit changes to support new cache eviction model in WebCore.
* WebView/WebPreferences.m:
(+[WebPreferences initialize]): Modified to reflect new API in WebCore.
* WebView/WebView.mm:
(+[WebView _initializeCacheSizesIfNecessary]): Slightly increased cache
size on low memory systems to avoid affecting the PLT for now.
win:
Reviewed by Dave Hyatt.
WebKit changes to support new cache eviction model in WebCore.
* WebPreferences.cpp:
(WebPreferences::initialize):
* WebView/WebPreferences.m: Modified to reflect new API in WebCore.
* WebView.cpp:
(WebView::initializeCacheSizesIfNecessary): Slightly increased cache
size on low memory systems to avoid affecting the PLT for now.
git-svn-id: http://svn.webkit.org/repository/webkit/trunk@25114 268f45cc-cd09-0410-ab3c-d52691b4dbfc
diff --git a/WebCore/loader/CachedResource.cpp b/WebCore/loader/CachedResource.cpp
index bae9d93..c0cb020 100644
--- a/WebCore/loader/CachedResource.cpp
+++ b/WebCore/loader/CachedResource.cpp
@@ -116,7 +116,7 @@
cache()->removeFromLiveResourcesSize(this);
cache()->removeFromLiveDecodedResourcesList(this);
allReferencesRemoved();
- cache()->pruneDeadResources();
+ cache()->prune();
}
}
@@ -146,7 +146,7 @@
cache()->removeFromLiveDecodedResourcesList(this);
// Update the cache's size totals.
- cache()->adjustSize(referenced(), delta, delta);
+ cache()->adjustSize(referenced(), delta);
}
}
@@ -173,7 +173,7 @@
cache()->insertInLRUList(this);
// Update the cache's size totals.
- cache()->adjustSize(referenced(), delta, 0);
+ cache()->adjustSize(referenced(), delta);
}
}
@@ -186,7 +186,7 @@
cache()->removeFromLiveDecodedResourcesList(this);
cache()->insertInLiveDecodedResourcesList(this);
}
- cache()->pruneLiveResources();
+ cache()->prune();
}
}
