Over a leisurely beer at our local pub, the Waggon and Horses, Phil Factor was holding forth on the esoteric, but strangely poetic, language of SQL Server internals, riddled as it is with ‘sleeping threads’, ‘stolen pages’, and ‘memory sweeps’. Generally, I remain immune to any twinge of interest in the bowels of SQL Server, reasoning that there are certain things that I don’t and shouldn’t need to know about SQL Server in order to use it successfully. Suddenly, however, my attention was grabbed by his mention of the ‘clock hands of the buffer cache’. Back at the office, I succumbed to a moment of weakness and opened up Google. He wasn’t lying.
SQL Server maintains various memory buffers, or caches. For example, the plan cache stores recently-used execution plans. The data cache in the buffer pool stores frequently-used pages, ensuring that they may be read from memory rather than via expensive physical disk reads. These memory stores are classic LRU (Least Recently Updated) buffers, meaning that, for example, the least frequently used pages in the data cache become candidates for eviction (after first writing the page to disk if it has changed since being read into the cache). SQL Server clearly needs some mechanism to track which pages are candidates for being cleared out of a given cache, when it is getting too large, and it is this mechanism that is somewhat more labyrinthine than I previously imagined.
Each page that is loaded into the cache has a counter, a miniature “wristwatch”, which records how recently it was last used. This wristwatch gets reset to “present time”, each time a page gets updated and then as the page ‘ages’ it clicks down towards zero, at which point the page can be removed from the cache. But what is SQL Server is suffering memory pressure and urgently needs to free up more space than is represented by zero-counter pages (or plans etc.)?
This is where our ‘clock hands’ come in. Each cache has associated with it a “memory clock”. Like most conventional clocks, it has two hands; one “external” clock hand, and one “internal“. Slava Oks is very particular in stressing that these names have “nothing to do with the equivalent types of memory pressure”. He’s right, but the names do, in that peculiar Microsoft tradition, seem designed to confuse.
The hands do relate to memory pressure; the cache “eviction policy” is determined by both global and local memory pressures on SQL Server. The “external” clock hand responds to global memory pressure, in other words pressure on SQL Server to reduce the size of its memory caches as a whole. Global memory pressure – which just to confuse things further seems sometimes to be referred to as physical memory pressure – can be either external (from the OS) or internal (from the process itself, e.g. due to limited virtual address space). The internal clock hand responds to local memory pressure, in other words the need to reduce the size of a single, specific cache. So, for example, if a particular cache, such as the plan cache, reaches a defined “pressure limit” the internal clock hand will start to turn and a memory sweep will be performed on that cache in order to remove plans from the memory store.
During each sweep of the hands, the usage counter on the cache entry is reduced in value, effectively moving its “last used” time to further in the past (in effect, setting back the wrist watch on the page a couple of hours) and increasing the likelihood that it can be aged out of the cache.
There is even a special Dynamic Management View, sys.dm_os_memory_cache_clock_hands, which allows you to interrogate the passage of the clock hands. Frequently turning hands equates to excessive memory pressure, which will lead to performance problems.
Two hours later, I emerged from this rather frightening journey into the heart of SQL Server memory management, fascinated but still unsure if I’d learned anything that I’d put to any practical use. However, I certainly began to agree that there is something almost Tolkeinian in the language of the deep recesses of SQL Server.