xfsqa 008 takes too long on UML...
PATCH 1/1 xfsqa 008 takes too long on UML...
With the recent change for reliability with 64k page size made to test 008,the file sizes got much larger. It appears that randholes actually reads the entire file, so this has slowed the test down by a factor of ten (all file sizes were increased by 10x). This means the test is now taking about 18 minutes to run on a UML session, and all the time is spent reading the files.
Instead, scale the file size based on the page size. We know how many holes we are trying to produce and the I/O size being used to produce them, so the size of the files can be finely tuned. Assuming a decent random distribution, if the number of blocks in the file is 4x the page size and the I/O size is page sized, this means that every I/O should generate a new hole and we'll only get a small amount of adjacent extents. This has passed over 10 times on ia64 w/ 64k page and another 15 times on UML with 4k page.
UML runtime is down from ~1000s to 5s, ia64 runtime is down from ~30s to 7s.
Cheers,
Dave.
Signature · 73 lines
Dave Chinner
Principal Engineer
SGI Australian Software Group
---
xfstests/008 | 13 +++++++------
xfstests/008.out | 10 +++++-----
2 files changed, 12 insertions(+), 11 deletions(-)
Index: xfs-cmds/xfstests/008
===================================================================
--- xfs-cmds.orig/xfstests/008 2008-04-28 15:57:02.000000000 +1000
+++ xfs-cmds/xfstests/008 2008-05-12 20:43:11.946325031 +1000
@@ -28,7 +28,8 @@ _cleanup()
_filter()
{
- sed -e "s/-b $pgsize/-b PGSIZE/g"
+ sed -e "s/-b $pgsize/-b PGSIZE/g" \
+ -e "s/-l .* -c/-l FSIZE -c/g"
}
# get standard environment, filters and checks
@@ -79,13 +80,13 @@ _setup_testdir
rm -f $here/$seq.out.full
-_do_test 1 50 "-l 50000000 -c 50 -b $pgsize"
-_do_test 2 100 "-l 100000000 -c 100 -b $pgsize"
-_do_test 3 100 "-l 100000000 -c 100 -b 512" # test partial pages
+_do_test 1 50 "-l `expr 200 \* $pgsize` -c 50 -b $pgsize"
+_do_test 2 100 "-l `expr 400 \* $pgsize` -c 100 -b $pgsize"
+_do_test 3 100 "-l `expr 400 \* $pgsize` -c 100 -b 512" # test partial pages
# rinse, lather, repeat for direct IO
-_do_test 4 50 "-d -l 50000000 -c 50 -b $pgsize"
-_do_test 5 100 "-d -l 100000000 -c 100 -b $pgsize"
+_do_test 4 50 "-d -l `expr 200 \* $pgsize` -c 50 -b $pgsize"
+_do_test 5 100 "-d -l `expr 400 \* $pgsize` -c 100 -b $pgsize"
# note: direct IO requires page aligned IO
# todo: realtime.
Index: xfs-cmds/xfstests/008.out
===================================================================
--- xfs-cmds.orig/xfstests/008.out 2008-04-28 15:57:02.000000000 +1000
+++ xfs-cmds/xfstests/008.out 2008-05-12 20:41:12.053827580 +1000
@@ -1,21 +1,21 @@
QA output created by 008
-randholes.1 : -l 50000000 -c 50 -b PGSIZE
+randholes.1 : -l FSIZE -c 50 -b PGSIZE
------------------------------------------
holes is in range
-randholes.2 : -l 100000000 -c 100 -b PGSIZE
+randholes.2 : -l FSIZE -c 100 -b PGSIZE
------------------------------------------
holes is in range
-randholes.3 : -l 100000000 -c 100 -b 512
+randholes.3 : -l FSIZE -c 100 -b 512
------------------------------------------
holes is in range
-randholes.4 : -d -l 50000000 -c 50 -b PGSIZE
+randholes.4 : -d -l FSIZE -c 50 -b PGSIZE
------------------------------------------
holes is in range
-randholes.5 : -d -l 100000000 -c 100 -b PGSIZE
+randholes.5 : -d -l FSIZE -c 100 -b PGSIZE
------------------------------------------
holes is in rangeSounds reasonable to me.
Might be nice to have the description as a comment in the test so it is easy to see the thought behind the calculations in the future.
--Tim
Quoted 24 lines · David Chinner
With the recent change for reliability with 64k page size made to test 008,the file sizes got much larger. It appears that randholes actually reads the entire file, so this has slowed the test down by a factor of ten (all file sizes were increased by 10x). This means the test is now taking about 18 minutes to run on a UML session, and all the time is spent reading the files.
Instead, scale the file size based on the page size. We know how many holes we are trying to produce and the I/O size being used to produce them, so the size of the files can be finely tuned. Assuming a decent random distribution, if the number of blocks in the file is 4x the page size and the I/O size is page sized, this means that every I/O should generate a new hole and we'll only get a small amount of adjacent extents. This has passed over 10 times on ia64 w/ 64k page and another 15 times on UML with 4k page.
UML runtime is down from ~1000s to 5s, ia64 runtime is down from ~30s to 7s.
Cheers,
Dave.