-
Type:
Bug
-
Resolution: Done
-
Priority:
Major - P3
-
None
-
Affects Version/s: WT2.6.1
-
Component/s: None
-
None
Hi!
We're using column-store table to store market data information for financial instruments.
Record number is composed of instrument ID (high 24-bits) and sequential record numbers (lower 40 bits).
Workload mostly consists of writes with rather rare reads.
The problem we see is that CPU utilization is quite high.
perf shows the following hot stacks:
10.05%
__wt_reconcile
__wt_evict
__wt_evict_lru_page
__evict_worker
start_thread
__clone
4.94%
__wt_page_inmem
__wt_cache_read
__wt_page_in_func
__wt_row_search
__wt_btcur_insert
__curfile_insert
...
4.83%
__wt_cell_unpack231
__wt_page_inmem
__wt_cache_read
__wt_page_in_func
__wt_row_search
__wt_btcur_insert
__curfile_insert
...
4.43%
__wt_reconcile
__wt_evict
__wt_evict_lru_page
__evict_server
start_thread
__clone
3.65%
__rec_cell_build_leaf_key
__wt_reconcile
__wt_evict
__wt_evict_lru_page
__evict_worker
start_thread
__clone
3.55%
memmove
__rec_cell_build_leaf_key
__wt_reconcile
__wt_evict
__wt_evict_lru_page
__evict_worker
start_thread
__clone
2.75%
__log_wrlsn_server
start_thread
__clone
2.38%
memmove
__rec_cell_build_leaf_key
__wt_reconcile
__wt_evict
__wt_evict_lru_page
__evict_server
start_thread
__clone
1.64%
__rec_txn_read
__wt_reconcile
__wt_evict
__wt_evict_lru_page
__evict_worker
start_thread
__clone
1.49%
__wt_row_search
__wt_btcur_insert
__curfile_insert
...
1.40%
__rec_cell_build_leaf_key
__wt_reconcile
__wt_evict
__wt_evict_lru_page
__evict_server
start_thread
__clone
1.36%
__wt_col_search
__wt_btcur_insert
__curfile_insert
...
1.20%
memcpy
__wt_reconcile
__wt_evict
__wt_evict_lru_page
__evict_worker
start_thread
__clone
(I omit application-specific frames at the bottom of the stacks)
Are there any settings we can try to tune?
Please advise.
Thanks!
- is depended on by
-
SERVER-19532 WiredTiger changes for MongoDB 3.1.7
-
- Closed
-