diff options
| -rw-r--r-- | libglusterfs/src/iobuf.c | 4 | ||||
| -rw-r--r-- | tests/basic/write-behind.t | 53 | ||||
| -rw-r--r-- | xlators/mgmt/glusterd/src/glusterd-volume-set.c | 12 | ||||
| -rw-r--r-- | xlators/performance/write-behind/src/write-behind.c | 14 |
4 files changed, 77 insertions, 6 deletions
diff --git a/libglusterfs/src/iobuf.c b/libglusterfs/src/iobuf.c index 17cd68fc206..fa3ac840c43 100644 --- a/libglusterfs/src/iobuf.c +++ b/libglusterfs/src/iobuf.c @@ -30,8 +30,8 @@ struct iobuf_init_config gf_iobuf_init_config[] = { {8 * 1024, 128}, {32 * 1024, 64}, {128 * 1024, 32}, - {256 * 1024, 8}, - {1 * 1024 * 1024, 2}, + {256 * 1024, 64}, + {1 * 1024 * 1024, 64}, }; int diff --git a/tests/basic/write-behind.t b/tests/basic/write-behind.t new file mode 100644 index 00000000000..edad59786af --- /dev/null +++ b/tests/basic/write-behind.t @@ -0,0 +1,53 @@ +#!/bin/bash +# + +. $(dirname $0)/../include.rc +. $(dirname $0)/../volume.rc + +function clear_stats { + > /var/lib/glusterfs/stats/glusterfs_d_backends_${V0}0.dump +} + +function got_expected_write_count { + expected_size=$1 + expected_value=$2 + grep aggr.write_${expected_size} "/var/lib/glusterd/stats/glusterfsd__d_backends_${V0}0.dump" | grep $expected_value + if [ $? == 0 ]; then + echo "Y"; + else + echo "N"; + fi +} + +cleanup; + +TEST glusterd +TEST pidof glusterd + +TEST $CLI volume create $V0 replica 3 $H0:$B0/${V0}{0,1,2} + +# These are needed for our tracking of write sizes +TEST $CLI volume set $V0 diagnostics.latency-measurement on +TEST $CLI volume set $V0 diagnostics.count-fop-hits on +TEST $CLI volume set $V0 diagnostics.stats-dump-interval 2 + +# Disable this in testing to get deterministic results +TEST $CLI volume set $V0 performance.write-behind-trickling-writes off + +TEST $CLI volume start $V0 + +sleep 2; + +TEST glusterfs -s $H0 --volfile-id $V0 $M0 + +# Write a 100MB file with a window-size 1MB, we should get 100 writes of 1MB each +TEST dd if=/dev/zero of=$M0/100mb_file bs=1M count=100 +EXPECT_WITHIN 5 "Y" got_expected_write_count "1mb" 100 + +TEST $CLI volume set $V0 performance.write-behind-window-size 512KB + +# Write a 100MB file with a window-size 512KB, we should get 200 writes of 512KB each +TEST dd if=/dev/zero of=$M0/100mb_file_2 bs=1M count=100 +EXPECT_WITHIN 5 "Y" got_expected_write_count "512kb" 200 + +cleanup; diff --git a/xlators/mgmt/glusterd/src/glusterd-volume-set.c b/xlators/mgmt/glusterd/src/glusterd-volume-set.c index ac484e246b1..e82339a537f 100644 --- a/xlators/mgmt/glusterd/src/glusterd-volume-set.c +++ b/xlators/mgmt/glusterd/src/glusterd-volume-set.c @@ -1560,6 +1560,18 @@ struct volopt_map_entry glusterd_volopt_map[] = { .op_version = 2, .flags = OPT_FLAG_CLIENT_OPT }, + { .key = "performance.write-behind-trickling-writes", + .voltype = "performance/write-behind", + .option = "trickling-writes", + .op_version = 2, + .flags = OPT_FLAG_CLIENT_OPT + }, + { .key = "performance.nfs.write-behind-trickling-writes", + .voltype = "performance/write-behind", + .option = "trickling-writes", + .op_version = 2, + .flags = OPT_FLAG_CLIENT_OPT + }, { .key = "performance.lazy-open", .voltype = "performance/open-behind", .option = "lazy-open", diff --git a/xlators/performance/write-behind/src/write-behind.c b/xlators/performance/write-behind/src/write-behind.c index 7f5719b1e48..bc59036ff88 100644 --- a/xlators/performance/write-behind/src/write-behind.c +++ b/xlators/performance/write-behind/src/write-behind.c @@ -169,6 +169,7 @@ typedef struct wb_request { typedef struct wb_conf { uint64_t aggregate_size; + uint64_t page_size; uint64_t window_size; gf_boolean_t flush_behind; gf_boolean_t trickling_writes; @@ -1207,18 +1208,21 @@ __wb_collapse_small_writes (wb_request_t *holder, wb_request_t *req) char *ptr = NULL; struct iobuf *iobuf = NULL; struct iobref *iobref = NULL; + struct wb_conf *conf = NULL; int ret = -1; ssize_t required_size = 0; size_t holder_len = 0; size_t req_len = 0; + conf = req->wb_inode->this->private; + if (!holder->iobref) { holder_len = iov_length (holder->stub->args.vector, holder->stub->args.count); req_len = iov_length (req->stub->args.vector, req->stub->args.count); - required_size = max ((THIS->ctx->page_size), + required_size = max ((conf->page_size), (holder_len + req_len)); iobuf = iobuf_get2 (req->wb_inode->this->ctx->iobuf_pool, required_size); @@ -1281,7 +1285,6 @@ __wb_preprocess_winds (wb_inode_t *wb_inode) wb_request_t *holder = NULL; wb_conf_t *conf = NULL; int ret = 0; - ssize_t page_size = 0; /* With asynchronous IO from a VM guest (as a file), there can be two sequential writes happening in two regions @@ -1292,7 +1295,6 @@ __wb_preprocess_winds (wb_inode_t *wb_inode) through the interleaved ops */ - page_size = wb_inode->this->ctx->page_size; conf = wb_inode->this->private; list_for_each_entry_safe (req, tmp, &wb_inode->todo, todo) { @@ -1343,7 +1345,7 @@ __wb_preprocess_winds (wb_inode_t *wb_inode) continue; } - space_left = page_size - holder->write_size; + space_left = wb_inode->window_conf - holder->write_size; if (space_left < req->write_size) { holder->ordering.go = 1; @@ -2471,6 +2473,9 @@ reconfigure (xlator_t *this, dict_t *options) GF_OPTION_RECONF ("cache-size", conf->window_size, options, size_uint64, out); + GF_OPTION_RECONF ("cache-size", conf->page_size, options, size_uint64, + out); + GF_OPTION_RECONF ("flush-behind", conf->flush_behind, options, bool, out); @@ -2522,6 +2527,7 @@ init (xlator_t *this) /* configure 'option window-size <size>' */ GF_OPTION_INIT ("cache-size", conf->window_size, size_uint64, out); + GF_OPTION_INIT ("cache-size", conf->page_size, size_uint64, out); if (!conf->window_size && conf->aggregate_size) { gf_msg (this->name, GF_LOG_WARNING, 0, |