Time slicing enables there to be more logical threads than physical threads. Each logical thread is serviced for a time slice—a short period of time defined by the operating system during which a thread can run before being preempted—by a physical thread. If a thread runs longer than a time slice, as most do, it relinquishes the physical thread until it gets another turn. This chapter details the costs incurred by time slicing.

The most obvious cost is the time for context switching between logical threads. Each context switch requires that the processor save all its registers for the previous logical thread that it was executing, and load its registers with information for the next logical thread it runs.

A subtler cost is cache cooling. Processors keep recently accessed data in cache memory, which is very fast, but also relatively small compared to main memory. When the processor runs out of cache memory, it has to evict items from cache and put them back into main memory. Typically, it chooses the least recently-used items in the cache. (The reality of set-associative caches is a bit more complicated, but this is not a cache primer.)

When a logical thread gets its time slice, as it references a piece of data for the first time, this data is pulled into cache, taking hundreds of cycles. If it is referenced frequently enough not to be evicted, each subsequent reference will find it in cache, and take only a few cycles. Such data is called hot in cache.

Time slicing undoes this because if Thread A finishes its time slice, and subsequently Thread B runs on the same physical thread, B will tend to evict data that was hot in cache for A, unless both threads need the data. When Thread A gets its next time slice, it will need to reload evicted data, at the cost of hundreds of cycles for each cache miss. Or worse yet, the next time slice for Thread A may be on a different physical thread that has a different cache altogether.

Another cost is lock preemption. This happens if a thread acquires a lock on a resource and its time slice runs out before it releases the lock. No matter how short a time the thread intended to hold the lock, it is now going to hold it for at least as long as it takes for its next turn at a time slice to come up. Any other threads waiting on the lock either busy-wait pointlessly or lose the rest of their time slice. The effect is called convoying because the threads end up “bumper to bumper” waiting for the preempted thread in front to resume driving.