Details
-
Type:
Task
-
Status: Closed
-
Priority:
Major - P3
-
Resolution: Fixed
-
Affects Version/s: None
-
Fix Version/s: WT2.7.0
-
Labels:
-
# Replies:5
-
Last comment by Customer:true
-
External issue ID:
Description
It would be nice to be able to flush dirty pages from cache without evicting them. Some workloads don't generate a lot of cache pressure, but do want to do checkpoints.
If the cache is allowed to have a lot of dirty content then checkpoints become more expensive.
It would be good if code:eviction_dirty_target is set below evicting_target we switched to a mode of only "cleaning" dirty pages, rather than evicting them.
Issue Links
- is depended on by
-
SERVER-18647
WiredTiger changes for MongoDB 3.0.4
-
- Closed
-
- is related to
-
WT-1788
Review eviction server algorithm
-
- Open
-
-
WT-1744
Throttle inserts when the cache has lots of dirty content
-
- Closed
-
- links to
Activity
Alex Gorrod, Michael Cahill, I think I have this working, I was wondering if you had any workloads you thought might be interesting?
There are two things to look out for:
1) Tracking statistics of how much of the cache is dirty.
2) Watching for pauses due to the cache getting stuck full during a checkpoint.
For 1) I'd try running an update intensive wtperf workload. Probably the update-btree.wtperf workload. You may want to increase the cache size, to allow for a reasonable steady state.
For 2) I'd add in checkpoints to the update-btree.wtperf workload and look for pauses if the cache is running dirty.
If neither of those workloads are heavy enough to get the cache full, you could try updating the number of worker threads to increase the system load.
Author:
{u'username': u'keithbostic', u'name': u'Keith Bostic', u'email': u'keith@wiredtiger.com'}Message: If the cache is large enough but there are too many dirty pages, don't
evict the pages, clean them and leave them in the cache; ref WT-1350.
Abstract out the test for when the cache requires eviction or to reduce
the number of dirty pages into a new function, __wt_cache_status().
Change the page eviction function, __wt_evict() to call that function
after reconciling each page, and leave the page in the cache if the
cache is currently large enough.
Branch: develop
https://github.com/wiredtiger/wiredtiger/commit/cff63f7abbe007246bec6e844477e98a39d297c9
Author:
{u'username': u'agorrod', u'name': u'Alex Gorrod', u'email': u'alexg@wiredtiger.com'}Message: Add an eviction_dirty_trigger configuration setting.
Allows applications to maintain a low proportion of dirty pages in
the cache, while allowing clean pages to utilize the cache.
While here, alter the minimum allowed dirty trigger and target to
be set as low as 5 percent. Which may be desired for huge caches.
refs SERVER-16737, WT-1350
e Please enter the commit message for your changes. Lines starting
Branch: develop
https://github.com/wiredtiger/wiredtiger/commit/e075e2c9444549e718a98a345fa967cc92c5da09
Alex, a question about this: isn't that what we do with the initial pass before each checkpoint to clear out dirty pages?
I guess this would keep checkpoints from taking so long, is that the goal?