Increase default maximum allowed readahead window from 128 KB to 512 KB.
This improves performance for some workloads (see below for details) where
ability to scale readahead window to larger sizes allows for better total
throughput while chances for regression are rather low given readahead
window size is dynamically computed based on observation (and thus it never
grows large for workloads with a random read pattern).
Note that the same tuning can be done using udev rules or by manually setting
the sysctl parameter however we believe the new value is a better default most
users will want to use. As a data point we carry this patch in SUSE kernels
for over 8 years.
Some data from the last evaluation of this patch (on 4.4-based kernel, I can
rerun those tests on a newer kernel but nothing has changed in the readahead
area since 4.4). The patch was evaluated on two machines
o a UMA machine, 8 cores and rotary storage
o A NUMA machine, 4 socket, 48 cores and SSD storage
Five basic tests were conducted;
1. paralleldd-single
paralleldd uses different instances of dd to access a single file and
write the contents to /dev/null. The performance of it depends on how
well readahead works for a single file. It's mostly sequential IO.
2. paralleldd-multi
Similar to test 1 except each instance of dd accesses a different file
so each instance of dd is accessing data sequentially but the timing
makes it look like random read IO.
3. pgbench-small
A standard init of pgbench and execution with a small data set
4. pgbench-large
A standard init of pgbench and execution with a large data set
5. bonnie++ with dataset sizes 2X RAM and in asyncronous mode
UMA paralleldd-single on ext3
4.4.0 4.4.0
vanilla readahead-v1r1
Amean Elapsd-1 5.42 ( 0.00%) 5.40 ( 0.50%)
Amean Elapsd-3 7.51 ( 0.00%) 5.54 ( 26.25%)
Amean Elapsd-5 7.15 ( 0.00%) 5.90 ( 17.46%)
Amean Elapsd-7 5.81 ( 0.00%) 5.61 ( 3.42%)
Amean Elapsd-8 6.05 ( 0.00%) 5.73 ( 5.36%)
Results speak for themselves, readahead is a major boost when there
are multiple readers of data. It's not displayed but system CPU
usage is overall. The IO stats support the results
4.4.0 4.4.0
vanillareadahead-v1r1
Mean sda-avgqusz 7.44 8.59
Mean sda-avgrqsz 279.77 722.52
Mean sda-await 31.95 48.82
Mean sda-r_await 3.32 11.58
Mean sda-w_await 127.51 119.60
Mean sda-svctm 1.47 3.46
Mean sda-rrqm 27.82 23.52
Mean sda-wrqm 4.52 5.00
It shows that the average request size is 2.5 times larger even
though the merging stats are similar. It's also interesting to
note that average wait times are higher but more IO is being
initiated per dd instance.
It's interesting to note that this is specific to ext3 and that xfs showed
a small regression with larger readahead.
UMA paralleldd-single on xfs
4.4.0 4.4.0
vanilla readahead-v1r1
Min Elapsd-1 6.91 ( 0.00%) 7.10 ( -2.75%)
Min Elapsd-3 6.77 ( 0.00%) 6.93 ( -2.36%)
Min Elapsd-5 6.82 ( 0.00%) 7.00 ( -2.64%)
Min Elapsd-7 6.84 ( 0.00%) 7.05 ( -3.07%)
Min Elapsd-8 7.02 ( 0.00%) 7.04 ( -0.28%)
Amean Elapsd-1 7.08 ( 0.00%) 7.20 ( -1.68%)
Amean Elapsd-3 7.03 ( 0.00%) 7.12 ( -1.40%)
Amean Elapsd-5 7.22 ( 0.00%) 7.38 ( -2.34%)
Amean Elapsd-7 7.07 ( 0.00%) 7.19 ( -1.75%)
Amean Elapsd-8 7.23 ( 0.00%) 7.23 ( -0.10%)
The IO stats are not displayed but show a similar ratio to ext3 and system
CPU usage is also lower. Hence, this slowdown is unexplained but may be
due to differences in XFS in the read path and how it locks even though
direct IO is not a factor. Tracing was not enabled to see what flags are
passed into xfs_ilock to see if the IO is all behind one lock but it's
one potential explanation.
UMA paralleldd-single on ext3
This showed nothing interesting as the test was too short-lived to draw
any conclusions. There was some difference in the kernels but it was
within the noise. The same applies for XFS.
UMA pgbench-small on ext3
This showed very little that was interesting. The database load time
was slower but by a very small margin. The actual transaction times
were highly variable and inconclusive.
NUMA pgbench-small on ext3
Load times are not reported but they completed 1.5% faster.
4.4.0 4.4.0
vanilla readahead-v1r1
Hmean 1 3000.54 ( 0.00%) 2895.28 ( -3.51%)
Hmean 8 20596.33 ( 0.00%) 19291.92 ( -6.33%)
Hmean 12 30760.68 ( 0.00%) 30019.58 ( -2.41%)
Hmean 24 74383.22 ( 0.00%) 73580.80 ( -1.08%)
Hmean 32 88377.30 ( 0.00%) 88928.70 ( 0.62%)
Hmean 48 88133.53 ( 0.00%) 96099.16 ( 9.04%)
Hmean 80 55981.37 ( 0.00%) 76886.10 ( 37.34%)
Hmean 112 74060.29 ( 0.00%) 87632.95 ( 18.33%)
Hmean 144 51331.50 ( 0.00%) 66135.77 ( 28.84%)
Hmean 172 44256.92 ( 0.00%) 63521.73 ( 43.53%)
Hmean 192 35942.74 ( 0.00%) 71121.35 ( 97.87%)
The impact here is substantial particularly for higher thread-counts.
It's interesting to note that there is an apparent regression for low
thread counts. In general, there was a high degree of variability
but the gains were all outside of the noise. In general, the io stats
did not show any particular pattern about request size as the workload
is mostly resident in memory. The real curiousity is that readahead
should have had little or no impact here as the data is mostly resident
in memory. Observing the transactions over time, there was a lot of
variability and the performance is likely dominated by whether the
data happened to be local or not. In itself, this test does not push
for inclusion of the patch due to the lack of IO but is included for
completeness.
UMA pgbench-small on xfs
Similar observations to ext3 on the load times. The transaction times
were stable but showed no significant performance difference.
UMA pgbench-large on ext3
Database load times were slightly faster (3.36%). The transaction times
were slower on average, more variable but still very close to the noise.
UMA pgbench-large on xfs
No significant difference on either database load times or transactions.
UMA bonnie on ext3
4.4.0 4.4.0
vanilla readahead-v1r1
Hmean SeqOut Char 81079.98 ( 0.00%) 81172.05 ( 0.11%)
Hmean SeqOut Block 104416.12 ( 0.00%) 104116.24 ( -0.29%)
Hmean SeqOut Rewrite 44153.34 ( 0.00%) 44596.23 ( 1.00%)
Hmean SeqIn Char 88144.56 ( 0.00%) 91702.67 ( 4.04%)
Hmean SeqIn Block 134581.06 ( 0.00%) 137245.71 ( 1.98%)
Hmean Random seeks 258.46 ( 0.00%) 280.82 ( 8.65%)
Hmean SeqCreate ops 2.25 ( 0.00%) 2.25 ( 0.00%)
Hmean SeqCreate read 2.25 ( 0.00%) 2.25 ( 0.00%)
Hmean SeqCreate del 911.29 ( 0.00%) 880.24 ( -3.41%)
Hmean RandCreate ops 2.25 ( 0.00%) 2.25 ( 0.00%)
Hmean RandCreate read 2.00 ( 0.00%) 2.25 ( 12.50%)
Hmean RandCreate del 911.89 ( 0.00%) 878.80 ( -3.63%)
The difference in headline performance figures is marginal and well within noise.
The system CPU usage tells a slightly different story
4.4.0 4.4.0
vanillareadahead-v1r1
User 1817.53 1798.89
System 499.40 420.65
Elapsed 10692.67 10588.08
As do the IO stats
4.4.0 4.4.0
vanillareadahead-v1r1
Mean sda-avgqusz 1079.16 1083.35
Mean sda-avgrqsz 807.95 1225.08
Mean sda-await 7308.06 9647.13
Mean sda-r_await 119.04 133.27
Mean sda-w_await 19106.20 20255.41
Mean sda-svctm 4.67 7.02
Mean sda-rrqm 1.80 0.99
Mean sda-wrqm 5597.12 5723.32
NUMA bonnie on ext3
bonnie
4.4.0 4.4.0
vanilla readahead-v1r1
Hmean SeqOut Char 58660.72 ( 0.00%) 58930.39 ( 0.46%)
Hmean SeqOut Block 253950.92 ( 0.00%) 261466.37 ( 2.96%)
Hmean SeqOut Rewrite 151960.60 ( 0.00%) 161300.48 ( 6.15%)
Hmean SeqIn Char 57015.41 ( 0.00%) 55699.16 ( -2.31%)
Hmean SeqIn Block 600448.14 ( 0.00%) 627565.09 ( 4.52%)
Hmean Random seeks 0.00 ( 0.00%) 0.00 ( 0.00%)
Hmean SeqCreate ops 1.00 ( 0.00%) 1.00 ( 0.00%)
Hmean SeqCreate read 3.00 ( 0.00%) 3.00 ( 0.00%)
Hmean SeqCreate del 90.91 ( 0.00%) 79.88 (-12.14%)
Hmean RandCreate ops 1.00 ( 0.00%) 1.50 ( 50.00%)
Hmean RandCreate read 3.00 ( 0.00%) 3.00 ( 0.00%)
Hmean RandCreate del 92.95 ( 0.00%) 93.97 ( 1.10%)
The impact is small but in line with the UMA machine in a number of details.
As before, the CPU usage is lower even if the iostats show very little
differences overall.
Overall, the headline performance figures are mostly improved or show
little difference. There is a small anomaly with XFS that indicates it may
not always win there due to other factors. There is also the possibility
that a mostly random read workload that was larger than memory with each
read spanning multiple pages but less than the max readahead window would
suffer but the probability is low as the readahead window should scale
properly. On balance, this is a win -- particularly on the large read
workloads.
Signed-off-by: Jan Kara <jack@xxxxxxx>
---
include/linux/mm.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/include/linux/mm.h b/include/linux/mm.h
index 00bad7793788..c50c6f442786 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -1991,7 +1991,7 @@ int write_one_page(struct page *page, int wait);
void task_dirty_inc(struct task_struct *tsk);
/* readahead.c */
-#define VM_MAX_READAHEAD 128 /* kbytes */
+#define VM_MAX_READAHEAD 512 /* kbytes */
#define VM_MIN_READAHEAD 16 /* kbytes (includes current page) */
int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
