Intel hyper-threading can boost performance, but beware of higher licensing fees 

Intel's new hyper-threading technology could increase processor performance and can be used with symmetric multiprocessing (SMP). Software licensing is the hidden danger.

Event
On 28 August 2001, Intel announced hyper-threading technology, which the company plans to use in its Xeon processor family in 2002. Intel's hyper-threading technology enables an individual 32-bit Xeon processor to appear as two logical or virtual processors to the operating system (OS) and applications it hosts. The technology duplicates key processor resources (e.g., registers) that enable a processor to maintain the context for a processing thread. By duplicating these resources inside a processor, a single processor can simultaneously execute instructions from two threads and maintain the architectural state of each thread.

Hyper-threading resembles symmetrical processing between two processors but does not duplicate all of the resources inside a single processor. Thus, hyper-threading is neither dual core (i.e., two compete processors in one processor die) nor dual processor (i.e., two processors bundled on the same package). Hyper-threading and dual processor technologies differ in performance. With hyper-threading, a multithreaded application should achieve up to a 30 percent performance improvement over a uniprocessor and a 60 percent to 80 percent improvement over a uniprocessor using dual processors (e.g., dual core or packaged). Intel expects to use hyper-threading technology in non-Xeon processors eventually.

Hyper-threading can be combined with conventional SMP processor configurations. Intel will use hyper-threading with dual processor and multiprocessor Xeon configurations. Combining hyper-threading with SMP brings a new level of processing efficiency to the individual processor. However, in an SMP environment, processor-aware applications must deploy tasks first at the processor level (SMP) and then at the hyper-threading level. Without the distinction between SMP and hyper-threading, processor-intensive tasks may be deployed to the same processor rather than to different processors.

To take advantage of hyper-threading, a system must host an SMP-aware OS and applications. For example, Windows 98 and Windows Me are not SMP-aware, but Windows NT, Windows 2000, and Windows XP are. Applications can be written as single-threaded (non-SMP-aware) or multithreaded (SMP-aware), but only multithreaded applications will enjoy a performance boost. Using an SMP-aware OS with a single-threaded application can provide a marginal performance improvement.

Because a hyper-threaded processor appears to software as two processors, software can easily confuse a dual-processor Xeon system with a 4-way Xeon system. Software must learn to distinguish between hyper-threading and SMP. Otherwise, enterprises could pay higher fees based on per-processor licensing. Intel claims to have worked through this issue with Microsoft, but other server-side software vendors must come on board as well.

Enterprises should consider hyper-threading as a performance boost for multithreaded server applications, but they should be cautious about software that uses per-processor licensing. Enterprises should discuss this issue with their software vendors and obtain a written commitment that licensing fees will not increase due to hyper-threading.

Gartner originally published this report on Sept. 5, 2001.
Analytical Source: John Enck, NT Strategies

"Linux is for people who hate Windows. BSD is for 
people that love unix."