|Publication number||US20060117200 A1|
|Application number||US 11/336,303|
|Publication date||Jun 1, 2006|
|Filing date||Jan 20, 2006|
|Priority date||Jul 15, 2003|
|Also published as||CN1577280A, CN100555227C, CN101320289A, CN101320289B, DE112004001320B3, US7389440, US7392414, US7788519, US20050050310, US20060117199, US20060123263, US20060123264, US20070198872, WO2005010737A2|
|Publication number||11336303, 336303, US 2006/0117200 A1, US 2006/117200 A1, US 20060117200 A1, US 20060117200A1, US 2006117200 A1, US 2006117200A1, US-A1-20060117200, US-A1-2006117200, US2006/0117200A1, US2006/117200A1, US20060117200 A1, US20060117200A1, US2006117200 A1, US2006117200A1|
|Inventors||Daniel Bailey, Todd Dutton, Tryggve Fossum|
|Original Assignee||Intel Corporation|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (3), Referenced by (3), Classifications (16), Legal Events (1)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This application is a division of co-pending U.S. patent application Ser. No. 10/621,228 filed Jul. 15, 2003 and entitled “A Method, System, and Apparatus for Improving Multi-Core Processor Performance,” and is related to three concurrently filed U.S. Patent Applications, Attorney Docket Nos. 042390.P16355D, 042390.P16355D3 and 042390.P16355D4, also entitled “A Method, System, and Apparatus for Improving Multi-Core Processor Performance.”
The present disclosure pertains to the field of power management. More particularly, the present disclosure pertains to a new method and apparatus for improving multi-core processor performance despite power constraints.
2. Description of Related Art
Power management schemes allow for reducing power consumption to achieve low power applications for various types of and systems and integrated devices, such as, servers, laptops, processors and desktops. Typically, software methods are employed for systems and integrated devices to support multiple power states for optimizing performance based at least in part on the Central Processing Unit (CPU) activity.
Present power management schemes either decrease voltage or frequency or both for reducing power consumption. However, this results in decreased overall performance. Also, some methods incorporate analog designs that have various challenges relating to loop stability for transient workloads, calibration, and tuning.
With the introduction of processors with multiple cores, power management becomes a major concern because of the increase in cores operating at high frequencies and voltages and need to adhere to various power constraints, such as, thermal limits, maximum current, and Vcc range.
The present invention is illustrated by way of example and not limitation in the Figures of the accompanying drawings.
The following description provides method and apparatus for improved multi-core processor performance despite power constraints. In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the present invention. It will be appreciated, however, by one skilled in the art that the invention may be practiced without such specific details. Those of ordinary skill in the art, with the included descriptions, will be able to implement appropriate logic circuits without undue experimentation.
As previously described, a problem exists for improving processor performance while adhering to power constraints. The present methods incorporate lowering the voltage or frequency at the expense of overall performance. In contrast, the claimed subject matter improves overall performance while adhering to power constraints. For example, a concept of “rationing the number of executing cores for a processor system” allows for increasing frequency as a result of disabling clocks to cores that are idle as they wait for a memory transaction to complete. For example, the claimed subject matter exploits the idle time period of processor cores by disabling the clocks to the core, that results in less power dissipation. Thus, a higher frequency can be utilized as a result of the decrease in power dissipation. In one embodiment, an appropriate executing core limit is calculated for the workload. Also, in the same embodiment, the number of executing cores are less than or equal to the number of available and ready threads. A thread is an independent set of instructions for a particular application.
In one embodiment, the claimed subject matter facilitates selecting a voltage/frequency operating point based on a prediction of the activity level of the threads running on all of the cores collectively. For example, TPC-C threads tend to be active 50-60% of the time, and spend 40-50% of their time idle, waiting for memory references to be completed. In such an environment, one would specify an executing core limit that would be, in one embodiment, 60% of the total number of cores on the die; if there were 8 cores, one would set the executing core limit to, in this case, five. One would then specify a voltage-frequency operating point that corresponds to having only five cores active and three cores inactive (low power state) at a time; this is a significantly higher operating frequency than one would specify if one was allowing all eight cores to be simultaneously active. The core rationing logic constrains the operations of the die, guaranteeing that no more than five cores (in this case) are active at any given moment. Statistics are gathered regarding the occupancy of the Waiting and Rationing queues (which will be discussed further in connection with
In the same embodiment, the state diagram illustrates a predetermined state machine for a processor core in a system. In this same embodiment, the state machine facilitates the “rationing of the cores” to improve processor performance as a result of disabling clocks to cores that are waiting for a memory transaction to complete.
In one embodiment, the state diagram has four defined states, such as, a Core Unassigned state 202, an Executing state 204, a Rationing FIFO Queue state 206, and a Waiting state 208. Initially, the Core Unassigned state is defined as follows: each core does not have an assigned thread. Subsequently, in the event that a core has a thread assigned to it, the claimed subject matter transitions to the Rationing FIFO Queue state 206. In one embodiment, FIFO is defined as a First In First Out.
Upon transitioning to the Rationing FIFO Queue state, a comparison between the number of executing cores and an executing core limit (ECL) is determined. In one embodiment, a processor or system specification determines the proper executing core limit in order to adhere to thermal power considerations. In one embodiment, the ECL is determined by a formula depicted later in the application. If the number of executing cores is less than ECL, the particular core transitions to the Executing state 204 if the core was the next one to be processed in the FIFO queue. Otherwise, the core remains in the Rationing FIFO queue 206.
Upon entering the Executing state, the core remains in this state unless an event occurs, such as, a memory reference and overheating event, and/or a fairness timeout. For example, a fairness timeout may be utilized to prevent a possible live lock state. In this context, a memory reference refers to a read or write operation to a particular memory address that does not reside in any cache memory coupled to the processor (“a miss in all levels of cache memory”). Therefore, an access to main memory is initiated.
If an event occurs as previously described, the core transitions to the Waiting state 208. Upon completion of the event, the core transitions to the Rationing FIFO queue state. This sequence of cycling between states 204, 206, and 208 occurs until the particular thread is completed. Upon completion of the thread, the core transitions to the Core Unassigned State.
However, the claimed subject matter is not limited to the four defined states in the state diagram. The claimed subject matter supports different amounts of states.
Also, other embodiments of state diagrams would allow multiple priority levels for cores, as well as having different waiting queues depending on the nature of the event that provoked exit from the executing state (memory wait, thermal wait, ACPI wait, etc).
Typically, a core executes a memory read or write operation and subsequently executes an operation that is dependent on that operation (for example, it makes use of the data returned by a memory read operation). Subsequently, the core “stalls” waiting for that memory operation to be completed. In such a case, it asserts a signal to the central core rationing logic indicating that it is stalled; this is the indication that it is eligible to be disabled by the core rationing logic. The core rationing logic responds to this signal by “napping” the core in question—it asserts a “nap” signal to the core, which causes the core to block instruction issue and then transition into a (cache-coherent) low power state. Furthermore, the core rationing logic puts an identifier for that core in the Waiting queue. When the memory operation completes, the core deasserts the “stall” signal; the core rationing logic responds to this by moving the identifier for that core from the Waiting queue to the Rationing queue. If the number of currently executing (not “napped”) cores is less than or equal to the Executing Core Limit, the core rationing logic removes the oldest identifier from the Rationing queue, and deasserts the “nap” signal to that core.
Analyzing the 50% memory reference duty cycle highlights the fact that the percentage time executing saturates at 50%. Thus, processing the memory references consumes half of the executing time when the ECL is equal to the number of available threads.
In one embodiment, a self optimization formula is utilized to determine the appropriate ECL. In the formula, N depicts the number of threads that have context: % E depicts the percentage executing time; and % M depicts the percentage memory reference time. The formula is:
int (N×(% E/(% E+% M)))
While certain exemplary embodiments have been described and shown in the accompanying drawings, it is to be understood that such embodiments are merely illustrative of and not restrictive on the broad invention, and that this invention not be limited to the specific constructions and arrangements shown and described, since various other modifications may occur to those ordinarily skilled in the art upon studying this disclosure.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US6804632 *||Dec 6, 2001||Oct 12, 2004||Intel Corporation||Distribution of processing activity across processing hardware based on power consumption considerations|
|US7093147 *||Apr 25, 2003||Aug 15, 2006||Hewlett-Packard Development Company, L.P.||Dynamically selecting processor cores for overall power efficiency|
|US20040128663 *||Dec 31, 2002||Jul 1, 2004||Efraim Rotem||Method and apparatus for thermally managed resource allocation|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US7757233 *||Jul 10, 2008||Jul 13, 2010||International Business Machines Corporation||Controlling a computer system having a processor including a plurality of cores|
|US8769323 *||Mar 7, 2013||Jul 1, 2014||Intel Corporation||Method, apparatus, and system for optimizing frequency and performance in a multidie microprocessor|
|US8806248||Sep 30, 2013||Aug 12, 2014||Intel Corporation||Method, apparatus, and system for optimizing frequency and performance in a multidie microprocessor|
|Cooperative Classification||G06F1/3296, Y02B60/144, G06F1/329, G06F1/324, G06F1/3228, Y02B60/1282, Y02B60/1285, Y02B60/1217, G06F1/3287|
|European Classification||G06F1/32P5V, G06F1/32P5T, G06F1/32P1D, G06F1/32P5S, G06F1/32P5F|
|Jan 20, 2006||AS||Assignment|
Owner name: INTEL CORPORATION, CALIFORNIA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:BAILEY, DANIEL W.;DUTTON, TODD;FOSSUM, TRYGGVE;REEL/FRAME:017506/0187;SIGNING DATES FROM 20030918 TO 20040127