To: Joe NYC who wrote (76027 ) 4/1/2002 2:16:50 PM From: Petz Read Replies (2) | Respond to Intel's Hyperthreading implementation is almost certainly penalized by the method of sharing the trace cache between the two threads. From the Intel document, 2 for 1: Intel's Hyper-Threading Makes Xeon Processors do Double Duty ,...The L1, L2 and L3 chaces do not keep track of separate threads, they simply hopld data used by all threads. How this sharing is done has not been detailed by Intel, although surely some part of the caches are shared and some allocated... Why do that, especially with the tiny L1 D-cache? Anyway, HTh certainly increases the memory footprint of any software at the cache level, so even a perfect sharing scheme could slow down execution, compared to running each thread sequentially in time slices.The trace cache... uses a different scheme to track the two threads. Tags associated with the decoded instructions use a one-bit switch to identify the thread they belong to. This necessary step prevents all instructions from one thread being flushed from the cache when one thread branches to a new location in the code: only the instructions for that thread are thrown away.
