System interconnect dynamic scaling by predicting I/O requirements

• US 9,324,031 B2
• Filed: 06/09/2014
• Issued: 04/26/2016
• Est. Priority Date: 01/06/2014
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Dreps, Daniel Mark, Ferraiolo, Frank D., Haridass, Anand, Jayaraman, Prasanna, Krolak, David John
• Primary Examiner(s): Tsai, Henry
• Assistant Examiner(s): ROCHE, JOHN B
• Application Number:
  US14/299,415
• Publication Number:
  US 20150192981A1
• Time in Patent Office: 687 Days
• US Class Current: 1/1
• CPC Class Codes: G06F 1/3206 Monitoring of events, devices or parameters that trigger a change in power modality
• CPC Class Codes: G06F 1/3234 Power saving characterised by the action undertaken
• CPC Class Codes: G06F 1/3278 Power saving in modem or I/O interface
• CPC Class Codes: G06F 2009/45579 I/O management, e.g. providing access to device drivers or storage
• CPC Class Codes: G06F 2009/45591 Monitoring or debugging support
• CPC Class Codes: G06F 9/45558 Hypervisor-specific management and integration aspects
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks
• CPC Class Codes: Y02D 10/00 Energy efficient computing, e.g. low power processors, power management or thermal management

Abstract

Interface management techniques provide reduced power consumption along with reducing heat and EMI generation in a computer system having multiple interconnected processing units. Physical link layers of external interfaces that interconnect the processing units of have dynamically adjustable bandwidth and the bandwidths are dynamically adjusted by predicting interface bandwidth requirements. An interface control method detects events other than I/O requests that occur in a processing unit that are indicators of potential future transactions on one of the external interfaces connected to the processing unit. The method predicts, from the detected events, that future transactions will likely occur on the interface, and in response, controls the dynamically adjustable bandwidth of physical link layer of the interface to accommodate the future transactions.

Claims

1. A method of operating a computer system comprising a plurality of processing units interconnected by external interfaces having corresponding physical link layers with dynamically adjustable bandwidth, the method comprising:

detecting multiple events other than I/O requests and bandwidth allocation requests occurring within a first one of the processing units that are indicative of potential future transactions on a first one of the external interfaces connected to the first processing unit that connects the first processing unit to another processing unit;

predicting, from the detected events, that future transactions will likely occur on the first external interface, wherein the predicting is independent of observations of actual traffic on the external interface and requests for allocation of bandwidth on the external interface;

responsive to predicting that the future transactions will likely occur, controlling the dynamically adjustable bandwidth of a first physical link layer corresponding to the first external interface to accommodate the future transactions by directly signaling a controller that controls the dynamically adjustable bandwidth of the physical link layer to increase the dynamically adjustable bandwidth of the first physical link layer; and

subsequent to the future transactions occurring, restoring the dynamically adjustable bandwidth of the first physical link layer to a lower value.

2. The method of claim 1, wherein the detecting comprises:

within the first processing unit, snooping an internal bus of the first processing unit to detect events indicative of the potential future transactions on the first external interface; and

counting the events detected by the snooping, wherein the predicting compares the count to one or more thresholds to determine a quantity of the future potential transactions that may occur, and wherein the controlling controls the dynamically adjustable bandwidth to accommodate the quantity of future potential transactions.

3. The method of claim 1, wherein the detecting is performed by a hypervisor executing within the computer system detecting a system level event predictive of the future transactions and wherein the predicting predicts the future transactions responsive to the hypervisor detecting the system level event.

4. The method of claim 3, wherein the first external interface couples the first processing unit to a remote memory, wherein the hypervisor detecting a system level event comprises the hypervisor detecting that an association between a thread executed by the first processing unit and the remote memory has become active.

5. The method of claim 1, wherein the detecting comprises:

capturing states of the first processing unit within a trace array;

reading state information that has been captured in the trace array; and

further using the state information to detect the events indicative of the potential future transactions.

6. The method of claim 1, wherein the predicting computes a profile of a predicted utilization of the first interface and wherein the controlling controls the dynamically adjustable bandwidth in conformity with the shape of the predicted utilization.

7. The method of claim 1, wherein, responsive to the predicting determining that the future transactions will likely occur, the controlling sets the dynamically adjustable bandwidth of the first physical link layer to a predetermined value during periodic intervals, and wherein responsive to the predicting determining that the future transactions will not occur, the controlling sets the dynamically adjustable bandwidth of the first physical link layer to a minimum value.

8. The method of claim 1, wherein the events are idle cycles of a logical link layer of the first interface internal to the first processing unit and coupled to the first physical link layer.

9. The method of claim 1, wherein the controlling is performed by a bus arbiter of a link layer of the first interface internal to the first processing unit and coupled to the first physical link layer.

10. The method of claim 1, wherein the controlling controls a width of the first physical link layer by setting a number of active signals use to transmit data in the first physical link layer to set the dynamically adjustable bandwidth.

11. The method of claim 1, wherein the controlling controls a frequency of the first physical link layer to set the dynamically adjustable bandwidth.