As the name suggests, per-CPU variables work by keeping a copy of the variable, the data item in question, assigned to each (live) CPU on the system. In effect, we get rid of the problem area for concurrency, the critical section, by avoiding the sharing of data between threads. With the per-CPU data technique, since every CPU refers to its very own copy of the data, a thread running on that processor can manipulate it without any worry of racing. (This is roughly analogous to local variables; as locals are on the private stack of each thread, they aren't shared between threads, thus there's no critical section and no need for locking.) Here, too, the need for locking is thus eliminated – making it a lock-free technology!
So, think of this: if you are running on a system with four live CPU cores, then a per-CPU variable on that system is essentially an array of four elements: element 0 represents the data value on the first CPU...
