Add a field trial for using lower sqlite cache sizes.

chrome/browser/history/text_database.cc

Index: chrome/browser/history/text_database.cc
===================================================================
--- chrome/browser/history/text_database.cc	(revision 106923)
+++ chrome/browser/history/text_database.cc	(working copy)
@@ -15,6 +15,7 @@
 #include "base/stringprintf.h"
 #include "base/utf_string_conversions.h"
 #include "chrome/browser/diagnostics/sqlite_diagnostics.h"
+#include "chrome/browser/history/history_field_trial.h"
 #include "sql/statement.h"
 #include "sql/transaction.h"
 
@@ -138,7 +139,13 @@
   // have 2-3 of these objects, each with their own 8MB, this adds up very fast.
   // We therefore reduce the size so when there are multiple objects, we're not
   // too big.
-  db_.set_cache_size(512);
+  if (HistoryFieldTrial::IsLowMemFieldTrial()) {
+    // FTS does merges of 16 pages at a time, so we need a working set size
+    // somewhat greater than that to prevent too much thrashing.
+    db_.set_cache_size(40);
+  } else {
+    db_.set_cache_size(512);

[jar (doing other things), 2011/10/24 18:40:55]

Rather than having exactly two choices, have you consider a ew choices in a
range, each in a field trial group?  The sizes you chose differ by more than an
order of magnitude, and I'm guessing you're curious about where the optimum
paramaters are.  I'm not suggesting 50... but perhaps 3-5??

[Scott Hess - ex-Googler, 2011/10/24 19:15:27]

Wait - from our chat I thought this was 32, which is why I suggested that it
might want to be raised a little.  I'm less confident going from 512 to 40.

[brettw, 2011/10/25 01:55:02]

How about I'll change this to 128. Depending on what those histograms look like,
we'll try some lower values in the future.

[Scott Hess - ex-Googler, 2011/10/26 21:44:59]

OK, seems reasonable.

One "thing" about having just two values is we won't know the shape of the
curve.  You will get some threshold information (is it too horrible to keep or
not), but you probably won't get any info about how 64 or 256 will work.  Of
course, it's entirely possible that 128, 256 and 512 would all have the same
results, in which case more than two also won't tell you anything!  [That's a
happy result, actually.]

+  }
 
   // Run the database in exclusive mode. Nobody else should be accessing the
   // database while we're running, and this will give somewhat improved perf.