GWicke has uploaded a new change for review. https://gerrit.wikimedia.org/r/300100
Change subject: Lower trickle_fsync_interval to 8mb ...................................................................... Lower trickle_fsync_interval to 8mb In production, we still see relatively bursty IO, with no or few writes followed by a burst every five seconds or so. This causes short spikes of iowait up into the dozens of percent, which in turn do impact p99 read latencies. Our current trickle fsync interval is 30mb, which is rather large. I think it's worth lowering this significantly, perhaps to 8mb. All of those values are significantly above typical SSD erase block sizes of 256k. Bug: T140825 Change-Id: I9d6d2517f9ea90f99cd52c94483064ced9074324 --- M modules/cassandra/templates/cassandra.yaml-2.2.erb 1 file changed, 1 insertion(+), 1 deletion(-) git pull ssh://gerrit.wikimedia.org:29418/operations/puppet refs/changes/00/300100/1 diff --git a/modules/cassandra/templates/cassandra.yaml-2.2.erb b/modules/cassandra/templates/cassandra.yaml-2.2.erb index 4f4d153..f05c5b7 100644 --- a/modules/cassandra/templates/cassandra.yaml-2.2.erb +++ b/modules/cassandra/templates/cassandra.yaml-2.2.erb @@ -403,7 +403,7 @@ # impacting read latencies. Almost always a good idea on SSDs; not # necessarily on platters. trickle_fsync: true -trickle_fsync_interval_in_kb: 30240 +trickle_fsync_interval_in_kb: 8192 # TCP port, for commands and data # For security reasons, you should not expose this port to the internet. Firewall it if needed. -- To view, visit https://gerrit.wikimedia.org/r/300100 To unsubscribe, visit https://gerrit.wikimedia.org/r/settings Gerrit-MessageType: newchange Gerrit-Change-Id: I9d6d2517f9ea90f99cd52c94483064ced9074324 Gerrit-PatchSet: 1 Gerrit-Project: operations/puppet Gerrit-Branch: production Gerrit-Owner: GWicke <gwi...@wikimedia.org> _______________________________________________ MediaWiki-commits mailing list MediaWiki-commits@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/mediawiki-commits