Data: Database article number more than 200W, every day around 3K increase.
Status: Lucene 2.9 + Pangu participle, read and write separation. The index file reaches more than 1G, and the list reads more and more slowly.
Requirements: Front page Implementation list seconds out, retrieve seconds out, enhance the user experience.
---------------------------------------------------------------------------------
Skipped Ms-sql partition, sub-Library, full-text index ... On MongoDB
---------------------------------------------------------------------------------
After: the Lucene 2.9 + Pangu participle + MongoDB + cache, the offline test effect is very good
After building the historical data, put it on the line (IBM System X3650 7979i09 Machine memory 16G) trial run, dumbfounded,
There's something wrong with intuition. But a wayward one, the database is exported, migrated to the IBM System X3650 M4 machine memory 24G, the CPU often jumps to 80%;
But the front-desk list is still fast and likes a MONGO. Other applications on the same machine are very slow.
Fun and anxiety together, open Mongostat, mongotop also found no abnormalities. A check was made of NUMA. I suddenly remembered the warning given when MongoDB was running.
The same day to the engine room, modify BIOS = Memory = Scoket Inter leave for Non-numa.
Restart the machine, the MongoDB process CPU is around 0.
--------------------------------------------------------------------------------------------------------
Microsoft NUMA Scenario: http://technet.microsoft.com/zh-cn/library/ms345345 (v=sql.105). aspx
MONGO deploy to Win2008 on a CPU that lasts 100%, change to Non-numa