Cache and eviction tuning

Cache size

The size of the cache is the single most important tuning knob for a WiredTiger application. Ideally the cache should be configured to be large enough to hold an application's working set.

The cache size for the database is normally configured by setting the cache_size configuration string when calling the wiredtiger_open function. The cache size can be adjusted after the open call with WT_CONNECTION::reconfigure.

An example of setting a cache size to 5GB:

    error_check(wiredtiger_open(
      home, NULL, "create,cache_size=5GB,log=(enabled,recover=on),statistics=(all)", &conn));

The effectiveness of the chosen cache size can be measured by reviewing the page eviction statistics for the database.

Cache resident objects

Objects can be created as cache resident - that is their contents will remain in cache and never be considered for the purposes of cache eviction. Cache residence can be configured with the WT_SESSION::create "cache_resident" configuration string. LSM tree objects do not support the "cache_resident" setting.

Configuring a cache resident object has several effects:

• Once the object's pages have been created or instantiated in memory no further I/O cost is ever paid for object access, minimizing potential latency.
• Cache resident objects can be accessed faster than objects tracked for potential eviction.
• If cache resident objects require a significant proportion of the configured cache size then non cache-resident objects can incur significantly higher I/O churn.
• If cache resident objects require more space than the configured cache size, then further operations will either return error or stall until space is made available by closing objects.

An example of configuring a cache-resident object:

    error_check(
      session->create(session, "table:mytable", "key_format=r,value_format=S,cache_resident=true"));

Eviction tuning

When an application approaches the maximum cache size, WiredTiger begins eviction to stop memory use from growing too large, approximating a least-recently-used algorithm.

WiredTiger provides several configuration options for tuning how pages are evicted from the cache. Different settings will improve performance depending on an application's particular workload. Customizing the eviction configuration settings can reduce latency spikes in application threads and can improve throughput in some applications.

WiredTiger eviction tuning options can be configured when first opening a database via wiredtiger_open, or changed after open with WT_CONNECTION::reconfigure.

The eviction_target configuration value (default 80%) is the level at which WiredTiger attempts to keep the overall cache usage. Eviction worker threads are active when the cache contains at least this much content, expressed as a percentage of the total cache size.

The eviction_trigger configuration value (default 95%) is the level at which application threads start to perform eviction. This will throttle application operations, increasing operation latency, usually resulting in the cache usage staying at this level when there is more cache pressure than eviction worker threads can handle in the background.

Operations will stall when the cache reaches 100% of the cache size. Application may want to change these settings from their defaults to either increase the range in which worker threads operate before application threads are throttled, or to use a larger proportion of RAM, if eviction worker threads have no difficulty handling the cache pressure generated by the application.

The eviction_dirty_target (default 5%) and eviction_dirty_trigger (default 20%) operate in a similar way to the overall targets, but only apply to dirty data in cache. In particular, application threads will be throttled if the percentage of dirty data reaches the eviction_dirty_trigger. Any page that has been modified since it was read from disk is considered dirty.

The dirty eviction settings control how much work checkpoints have to do in the worst case, and also limit how much memory fragmentation is likely for memory allocations related to the cache. Most memory fragmentation is caused by workloads that generate a mix of updates (small allocations) with cache misses (large allocations). Limiting the percentage of cache that can be dirty limits the worst case fragmentation to the approximately the same level.

    /*
     * Configure eviction to begin at 90% full, and run until the cache is only 75% dirty.
     */
    error_check(wiredtiger_open(home, NULL,
      "create,eviction_trigger=90,eviction_dirty_target=75,eviction_dirty_trigger=90", &conn));

By default, WiredTiger cache eviction is handled by a single, separate thread. In a large, busy cache, a single thread will be insufficient (especially when the eviction thread must wait for I/O). The "eviction=(threads_min)" and "eviction=(threads_max)" configuration values can be used to configure the minimum and maximum number of additional threads WiredTiger will create to keep up with the application eviction load. Finally, if the Wiredtiger eviction threads are unable to keep up with application demand for cache space, application threads will be tasked with eviction as well, potentially resulting in latency spikes.

    /* Configure up to four eviction threads */
    error_check(
      wiredtiger_open(home, NULL, "create,eviction_trigger=90,eviction=(threads_max=4)", &conn));