The generic writeback routines are departing from congestion_wait()
in preference of get_request_wait(), aka. waiting on the block queues.
Introduce the missing writeback wait queue for NFS, otherwise its
writeback pages will grow greedily, exhausting all PG_dirty pages.
Tests show that it can effectively reduce stalls in the disk-network
pipeline, improve performance and reduce delays.
The test cases are basically
for run in 1 2 3
for nr_dd in 1 10 100
for dirty_thresh in 10M 100M 1000M 2G
start $nr_dd dd's writing to a 1-disk mem=12G NFS server
During all tests, nfs_congestion_kb is set to 1/8 dirty_thresh.
3.2.0-rc1 3.2.0-rc1-ioless-full+
(w/o patch) (w/ patch)
----------- ------------------------
20.66 +136.7% 48.90 thresh=1000M/nfs-100dd-1
20.82 +147.5% 51.52 thresh=1000M/nfs-100dd-2
20.57 +129.8% 47.26 thresh=1000M/nfs-100dd-3
35.96 +96.5% 70.67 thresh=1000M/nfs-10dd-1
37.47 +89.1% 70.85 thresh=1000M/nfs-10dd-2
34.55 +106.1% 71.21 thresh=1000M/nfs-10dd-3
58.24 +28.2% 74.63 thresh=1000M/nfs-1dd-1
59.83 +18.6% 70.93 thresh=1000M/nfs-1dd-2
58.30 +31.4% 76.61 thresh=1000M/nfs-1dd-3
23.69 -10.0% 21.33 thresh=100M/nfs-100dd-1
23.59 -1.7% 23.19 thresh=100M/nfs-100dd-2
23.94 -1.0% 23.70 thresh=100M/nfs-100dd-3
27.06 -0.0% 27.06 thresh=100M/nfs-10dd-1
25.43 +4.8% 26.66 thresh=100M/nfs-10dd-2
27.21 -0.8% 26.99 thresh=100M/nfs-10dd-3
53.82 +4.4% 56.17 thresh=100M/nfs-1dd-1
55.80 +4.2% 58.12 thresh=100M/nfs-1dd-2
55.75 +2.9% 57.37 thresh=100M/nfs-1dd-3
15.47 +1.3% 15.68 thresh=10M/nfs-10dd-1
16.09 -3.5% 15.53 thresh=10M/nfs-10dd-2
15.09 -0.9% 14.96 thresh=10M/nfs-10dd-3
26.65 +13.0% 30.10 thresh=10M/nfs-1dd-1
25.09 +7.7% 27.02 thresh=10M/nfs-1dd-2
27.16 +3.3% 28.06 thresh=10M/nfs-1dd-3
27.51 +78.6% 49.11 thresh=2G/nfs-100dd-1
22.46 +131.6% 52.01 thresh=2G/nfs-100dd-2
12.95 +289.8% 50.50 thresh=2G/nfs-100dd-3
42.28 +81.0% 76.52 thresh=2G/nfs-10dd-1
40.33 +78.8% 72.10 thresh=2G/nfs-10dd-2
42.52 +67.6% 71.27 thresh=2G/nfs-10dd-3
62.27 +34.6% 83.84 thresh=2G/nfs-1dd-1
60.10 +35.6% 81.48 thresh=2G/nfs-1dd-2
66.29 +17.5% 77.88 thresh=2G/nfs-1dd-3
1164.97 +41.6% 1649.19 TOTAL write_bw
The local queue time for WRITE RPCs could be reduced by several orders!
3.2.0-rc1 3.2.0-rc1-ioless-full+
----------- ------------------------
90226.82 -99.9% 92.07 thresh=1000M/nfs-100dd-1
88904.27 -99.9% 80.21 thresh=1000M/nfs-100dd-2
97436.73 -99.9% 87.32 thresh=1000M/nfs-100dd-3
62167.19 -99.3% 444.25 thresh=1000M/nfs-10dd-1
64150.34 -99.2% 539.38 thresh=1000M/nfs-10dd-2
78675.54 -99.3% 540.27 thresh=1000M/nfs-10dd-3
5372.84 +57.8% 8477.45 thresh=1000M/nfs-1dd-1
10245.66 -51.2% 4995.71 thresh=1000M/nfs-1dd-2
4744.06 +109.1% 9919.55 thresh=1000M/nfs-1dd-3
1727.29 -9.6% 1562.16 thresh=100M/nfs-100dd-1
2183.49 +4.4% 2280.21 thresh=100M/nfs-100dd-2
2201.49 +3.7% 2281.92 thresh=100M/nfs-100dd-3
6213.73 +19.9% 7448.13 thresh=100M/nfs-10dd-1
8127.01 +3.2% 8387.06 thresh=100M/nfs-10dd-2
7255.35 +4.4% 7571.11 thresh=100M/nfs-10dd-3
1144.67 +20.4% 1378.01 thresh=100M/nfs-1dd-1
1010.02 +19.0% 1202.22 thresh=100M/nfs-1dd-2
906.33 +15.8% 1049.76 thresh=100M/nfs-1dd-3
642.82 +17.3% 753.80 thresh=10M/nfs-10dd-1
766.82 -21.7% 600.18 thresh=10M/nfs-10dd-2
575.95 +16.5% 670.85 thresh=10M/nfs-10dd-3
21.91 +71.0% 37.47 thresh=10M/nfs-1dd-1
16.70 +105.3% 34.29 thresh=10M/nfs-1dd-2
19.05 -71.3% 5.47 thresh=10M/nfs-1dd-3
123877.11 -99.0% 1187.27 thresh=2G/nfs-100dd-1
122353.65 -98.8% 1505.84 thresh=2G/nfs-100dd-2
101140.82 -98.4% 1641.03 thresh=2G/nfs-100dd-3
78248.51 -98.9% 892.00 thresh=2G/nfs-10dd-1
84589.42 -98.6% 1212.17 thresh=2G/nfs-10dd-2
89684.95 -99.4% 495.28 thresh=2G/nfs-10dd-3
10405.39 -6.9% 9684.57 thresh=2G/nfs-1dd-1
16151.86 -48.5% 8316.69 thresh=2G/nfs-1dd-2
16119.17 -49.0% 8214.84 thresh=2G/nfs-1dd-3
1177306.98 -92.1% 93588.50 TOTAL nfs_write_queue_time
The average COMMIT size is not impacted too much.
3.2.0-rc1 3.2.0-rc1-ioless-full+
----------- ------------------------
5.56 +44.9% 8.06 thresh=1000M/nfs-100dd-1
4.14 +109.1% 8.67 thresh=1000M/nfs-100dd-2
5.46 +16.3% 6.35 thresh=1000M/nfs-100dd-3
52.04 -8.4% 47.70 thresh=1000M/nfs-10dd-1
52.33 -13.8% 45.09 thresh=1000M/nfs-10dd-2
51.72 -9.2% 46.98 thresh=1000M/nfs-10dd-3
484.63 -8.6% 443.16 thresh=1000M/nfs-1dd-1
492.42 -8.2% 452.26 thresh=1000M/nfs-1dd-2
493.13 -11.4% 437.15 thresh=1000M/nfs-1dd-3
32.52 -72.9% 8.80 thresh=100M/nfs-100dd-1
36.15 +26.1% 45.58 thresh=100M/nfs-100dd-2
38.33 +0.4% 38.49 thresh=100M/nfs-100dd-3
5.67 +0.5% 5.69 thresh=100M/nfs-10dd-1
5.74 -1.1% 5.68 thresh=100M/nfs-10dd-2
5.69 +0.9% 5.74 thresh=100M/nfs-10dd-3
44.91 -1.0% 44.45 thresh=100M/nfs-1dd-1
44.22 -0.6% 43.96 thresh=100M/nfs-1dd-2
44.18 +0.2% 44.28 thresh=100M/nfs-1dd-3
1.42 +1.1% 1.43 thresh=10M/nfs-10dd-1
1.48 +0.3% 1.48 thresh=10M/nfs-10dd-2
1.43 -1.0% 1.42 thresh=10M/nfs-10dd-3
5.51 -6.8% 5.14 thresh=10M/nfs-1dd-1
5.91 -8.1% 5.43 thresh=10M/nfs-1dd-2
5.44 +3.0% 5.61 thresh=10M/nfs-1dd-3
8.80 +6.6% 9.38 thresh=2G/nfs-100dd-1
8.51 +65.2% 14.06 thresh=2G/nfs-100dd-2
15.28 -13.2% 13.27 thresh=2G/nfs-100dd-3
105.12 -24.9% 78.99 thresh=2G/nfs-10dd-1
101.90 -9.1% 92.60 thresh=2G/nfs-10dd-2
106.24 -29.7% 74.65 thresh=2G/nfs-10dd-3
909.85 +0.4% 913.68 thresh=2G/nfs-1dd-1
1030.45 -18.3% 841.68 thresh=2G/nfs-1dd-2
1016.56 -11.6% 898.36 thresh=2G/nfs-1dd-3
5222.74 -10.1% 4695.25 TOTAL nfs_commit_size
And here is the list of overall numbers.
3.2.0-rc1 3.2.0-rc1-ioless-full+
----------- ------------------------
1164.97 +41.6% 1649.19 TOTAL write_bw
54799.00 +25.0% 68500.00 TOTAL nfs_nr_commits
3543263.00 -3.3% 3425418.00 TOTAL nfs_nr_writes
5222.74 -10.1% 4695.25 TOTAL nfs_commit_size
7.62 +89.2% 14.42 TOTAL nfs_write_size
1177306.98 -92.1% 93588.50 TOTAL nfs_write_queue_time
5977.02 -16.0% 5019.34 TOTAL nfs_write_rtt_time
1183360.15 -91.7% 98645.74 TOTAL nfs_write_execute_time
51186.59 -62.5% 19170.98 TOTAL nfs_commit_queue_time
81801.14 +3.6% 84735.19 TOTAL nfs_commit_rtt_time
133015.32 -21.9% 103926.05 TOTAL nfs_commit_execute_time
Feng: do more coarse grained throttle on each ->writepages rather than
on each page, for better performance and avoid throttled-before-send-rpc
deadlock
Signed-off-by: Feng Tang <feng.tang@xxxxxxxxx>
Signed-off-by: Wu Fengguang <fengguang.wu@xxxxxxxxx>
---
fs/nfs/client.c | 2
fs/nfs/write.c | 84 +++++++++++++++++++++++++++++++-----
include/linux/nfs_fs_sb.h | 1
3 files changed, 77 insertions(+), 10 deletions(-)
--- linux-next.orig/fs/nfs/write.c 2011-10-20 23:08:17.000000000 +0800
+++ linux-next/fs/nfs/write.c 2011-10-20 23:45:59.000000000 +0800
@@ -190,11 +190,64 @@ static int wb_priority(struct writeback_
* NFS congestion control
*/
+#define NFS_WAIT_PAGES (1024L >> (PAGE_SHIFT - 10))
int nfs_congestion_kb;
-#define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
-#define NFS_CONGESTION_OFF_THRESH \
- (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
+/*
+ * SYNC requests will block on (2*limit) and wakeup on (2*limit-NFS_WAIT_PAGES)
+ * ASYNC requests will block on (limit) and wakeup on (limit - NFS_WAIT_PAGES)
+ * In this way SYNC writes will never be blocked by ASYNC ones.
+ */
+
+static void nfs_set_congested(long nr, struct backing_dev_info *bdi)
+{
+ long limit = nfs_congestion_kb >> (PAGE_SHIFT - 10);
+
+ if (nr > limit && !test_bit(BDI_async_congested, &bdi->state))
+ set_bdi_congested(bdi, BLK_RW_ASYNC);
+ else if (nr > 2 * limit && !test_bit(BDI_sync_congested, &bdi->state))
+ set_bdi_congested(bdi, BLK_RW_SYNC);
+}
+
+static void nfs_wait_congested(int is_sync,
+ struct backing_dev_info *bdi,
+ wait_queue_head_t *wqh)
+{
+ int waitbit = is_sync ? BDI_sync_congested : BDI_async_congested;
+ DEFINE_WAIT(wait);
+
+ if (!test_bit(waitbit, &bdi->state))
+ return;
+
+ for (;;) {
+ prepare_to_wait(&wqh[is_sync], &wait, TASK_UNINTERRUPTIBLE);
+ if (!test_bit(waitbit, &bdi->state))
+ break;
+
+ io_schedule();
+ }
+ finish_wait(&wqh[is_sync], &wait);
+}
+
+static void nfs_wakeup_congested(long nr,
+ struct backing_dev_info *bdi,
+ wait_queue_head_t *wqh)
+{
+ long limit = nfs_congestion_kb >> (PAGE_SHIFT - 10);
+
+ if (nr < 2 * limit - min(limit / 8, NFS_WAIT_PAGES)) {
+ if (test_bit(BDI_sync_congested, &bdi->state))
+ clear_bdi_congested(bdi, BLK_RW_SYNC);
+ if (waitqueue_active(&wqh[BLK_RW_SYNC]))
+ wake_up(&wqh[BLK_RW_SYNC]);
+ }
+ if (nr < limit - min(limit / 8, NFS_WAIT_PAGES)) {
+ if (test_bit(BDI_async_congested, &bdi->state))
+ clear_bdi_congested(bdi, BLK_RW_ASYNC);
+ if (waitqueue_active(&wqh[BLK_RW_ASYNC]))
+ wake_up(&wqh[BLK_RW_ASYNC]);
+ }
+}
static int nfs_set_page_writeback(struct page *page)
{
@@ -205,11 +258,8 @@ static int nfs_set_page_writeback(struct
struct nfs_server *nfss = NFS_SERVER(inode);
page_cache_get(page);
- if (atomic_long_inc_return(&nfss->writeback) >
- NFS_CONGESTION_ON_THRESH) {
- set_bdi_congested(&nfss->backing_dev_info,
- BLK_RW_ASYNC);
- }
+ nfs_set_congested(atomic_long_inc_return(&nfss->writeback),
+ &nfss->backing_dev_info);
}
return ret;
}
@@ -221,8 +271,10 @@ static void nfs_end_page_writeback(struc
end_page_writeback(page);
page_cache_release(page);
- if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
- clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
+
+ nfs_wakeup_congested(atomic_long_dec_return(&nfss->writeback),
+ &nfss->backing_dev_info,
+ nfss->writeback_wait);
}
static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
@@ -323,10 +375,17 @@ static int nfs_writepage_locked(struct p
int nfs_writepage(struct page *page, struct writeback_control *wbc)
{
+ struct inode *inode = page->mapping->host;
+ struct nfs_server *nfss = NFS_SERVER(inode);
int ret;
ret = nfs_writepage_locked(page, wbc);
unlock_page(page);
+
+ nfs_wait_congested(wbc->sync_mode == WB_SYNC_ALL,
+ &nfss->backing_dev_info,
+ nfss->writeback_wait);
+
return ret;
}
@@ -342,6 +401,7 @@ static int nfs_writepages_callback(struc
int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
{
struct inode *inode = mapping->host;
+ struct nfs_server *nfss = NFS_SERVER(inode);
unsigned long *bitlock = &NFS_I(inode)->flags;
struct nfs_pageio_descriptor pgio;
int err;
@@ -358,6 +418,10 @@ int nfs_writepages(struct address_space
err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
nfs_pageio_complete(&pgio);
+ nfs_wait_congested(wbc->sync_mode == WB_SYNC_ALL,
+ &nfss->backing_dev_info,
+ nfss->writeback_wait);
+
clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
smp_mb__after_clear_bit();
wake_up_bit(bitlock, NFS_INO_FLUSHING);
--- linux-next.orig/include/linux/nfs_fs_sb.h 2011-10-20 23:08:17.000000000 +0800
+++ linux-next/include/linux/nfs_fs_sb.h 2011-10-20 23:45:12.000000000 +0800
@@ -102,6 +102,7 @@ struct nfs_server {
struct nfs_iostats __percpu *io_stats; /* I/O statistics */
struct backing_dev_info backing_dev_info;
atomic_long_t writeback; /* number of writeback pages */
+ wait_queue_head_t writeback_wait[2];
int flags; /* various flags */
unsigned int caps; /* server capabilities */
unsigned int rsize; /* read size */
--- linux-next.orig/fs/nfs/client.c 2011-10-20 23:08:17.000000000 +0800
+++ linux-next/fs/nfs/client.c 2011-10-20 23:45:12.000000000 +0800
@@ -1066,6 +1066,8 @@ static struct nfs_server *nfs_alloc_serv
INIT_LIST_HEAD(&server->layouts);
atomic_set(&server->active, 0);
+ init_waitqueue_head(&server->writeback_wait[BLK_RW_SYNC]);
+ init_waitqueue_head(&server->writeback_wait[BLK_RW_ASYNC]);
server->io_stats = nfs_alloc_iostats();
if (!server->io_stats) {
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
