LU-410 Revert LU-15 slow IO with read intense application

This reverts commit 12aae18c0ec495f459a1026ad2e86bbf0232b0fe.
This patch indirectly causes a performance drop with SRP.

Change-Id: If1af51ac8c2d45bc4d5cc5502dc59501678baaec
Signed-off-by: Johann Lombardi <johann@whamcloud.com>

diff --git a/lustre/ChangeLog b/lustre/ChangeLog
index 0c7ce5e..f67825e 100644 (file)
--- a/lustre/ChangeLog
+++ b/lustre/ChangeLog
@@ -41,13 +41,6 @@ Details    : The ldlm hash handling in Lustre 2.x is incompatible with the one
 
 Severity   : normal
 Jira       : LU-15
-Description: Reduce cache pressure on OSS
-Details    : Don't keep pages for objects >8MB in cache to alleviate memory
-            pressure on OSSs. This is done by changing the default value
-            of readcache_max_filesize to 8MB.
-
-Severity   : normal
-Jira       : LU-15
 Description: slow IO with read-intense application
 Details    : Align the readahead extent by 1M after when it is trimed by
             ra_max_pages.

diff --git a/lustre/obdfilter/filter_internal.h b/lustre/obdfilter/filter_internal.h
index 98190ce..ff7373f 100644 (file)
--- a/lustre/obdfilter/filter_internal.h
+++ b/lustre/obdfilter/filter_internal.h
@@ -113,8 +113,7 @@ enum {
 };
 
 //#define FILTER_MAX_CACHE_SIZE (32 * 1024 * 1024) /* was OBD_OBJECT_EOF */
-//#define FILTER_MAX_CACHE_SIZE OBD_OBJECT_EOF
-#define FILTER_MAX_CACHE_SIZE (8 * 1024 * 1024) /* was OBD_OBJECT_EOF */
+#define FILTER_MAX_CACHE_SIZE OBD_OBJECT_EOF
 
 /* We have to pass a 'created' array to fsfilt_map_inode_pages() which we
  * then ignore.  So we pre-allocate one that everyone can use... */