[IGNITE-10284] performance drop on PME time during 1 server node left

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Assignee: Unassigned
• Reporter: Max Shonichev
• Created: 15/Nov/18 20:24
• Updated: 19/Jun/19 11:21

Description

7 server nodes (1 node per host), 120 caches, G1GC, 32768 partitions 1 client tx loading

1. start cluster, fix baseline, preload data
2. start tx transfer task
3. kill 1 server node
4. wait for topology snapshot change, measure PME time
5. restart killed node
6. wait for topology snapshot change, measure PME time

for 1 server LEAVE, we see 7%-10% degradation, BUG

for 1 server JOIN, we see performance degradation on first exchange (evt=NODE_LEFT) and performance increase on second exchange (evt=CacheAffinityChange), overall 40-60% increase.

Issue Links

• relates to: IGNITE-9913 Prevent data updates blocking in case of backup BLT server node leave