Method and system for predicting communication delays of detailed application workloads
First Claim
1. A method for generating a delay model in a networked system under a defined workload, the method comprising:
- generating, for a first active message, a routing structure based upon a network configuration and a source network node and destination network node of the active message;
creating a contention structure created by summing together routing structure elements for active message events;
first calculating, for the first active message, an available bandwidth for use by the message at a path between network nodes utilized by the active message, the available bandwidth being a function of a level of contention between the first active message and at least a second active message on the path, the level of contention being determined in accordance with the contention structure and the routing structure for the first active message; and
second calculating, for the first active message, based upon the available bandwidth, a modeled communication delay to communicate at least a portion of the first active message.
2 Assignments
0 Petitions
Accused Products
Abstract
A method and system for generating a communication model are disclosed. The communication model is used to provide performance information on the use of communication networks by application code. The expected utilization of resources is important to when high performance is to be achieved in both cluster computing environments and information power GRIDs.
The disclosed method and system is designed to generate a model of an underlying network topology using detailed workload information, which encapsulates the expected computation/communication requirements of an application. Such workload information may be obtained from a number of different sources. Furthermore, the disclosed method provides detailed information of the expected use of a network system unlike traditional communication models that rely on statistical information and provide general performance information.
The disclosed method is a hybrid model containing statistical model information as well as steps for simulating the main stages that change the state of the communication network traffic. As such, the evaluation time is orders of magnitude quicker than traditional methods, while providing predictions, which are of reasonable accuracy. The characteristics of different networks can be easily incorporated into the model, and thus the model can be used in a variety of situations. The output of the model is the expected delay of a set of communications, which can be further, expanded into a suitable trace format and visualized for further investigation. In evaluation tests, the disclosed method and system provides predictions which are more accurate than simple regression models while requiring seconds of CPU processing time.
105 Citations
31 Claims
-
1. A method for generating a delay model in a networked system under a defined workload, the method comprising:
-
generating, for a first active message, a routing structure based upon a network configuration and a source network node and destination network node of the active message;
creating a contention structure created by summing together routing structure elements for active message events;
first calculating, for the first active message, an available bandwidth for use by the message at a path between network nodes utilized by the active message, the available bandwidth being a function of a level of contention between the first active message and at least a second active message on the path, the level of contention being determined in accordance with the contention structure and the routing structure for the first active message; and
second calculating, for the first active message, based upon the available bandwidth, a modeled communication delay to communicate at least a portion of the first active message. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10)
-
-
11. A computer system including executable program code for generating a delay model in a networked system under a defined workload, the computer system comprising:
-
an input stage that receives a workload description and renders event sequences corresponding to the workload description; and
an evaluation stage that receives the event sequences and renders timing information representing execution of the event sequences in a distributed processing network configuration, the evaluation stage comprising executable program instructions for;
generating for a first active message, a routing structure based upon a network configuration and a source network node and destination network node of the first active message;
creating a contention structure created by summing together routing structure elements for active message events;
first calculating, for the first active message, an available bandwidth for use by the message at a path between network nodes utilized by the first active message, the available bandwidth being a function of a level of contention between the first active message and at least a second active message on the path, the level of contention being determined in accordance with the contention structure and the routing structure for the first active message; and
second calculating, for the first active message, based upon the available bandwidth, a modeled communication delay to communicate at least a portion of the first active message. - View Dependent Claims (12, 13, 14, 15, 16, 17, 18, 19, 20)
-
-
21. A computer-readable medium having computer executable instructions for performing a set of steps to generate a delay model in a networked system under a defined workload, the steps including:
-
generating, for a first active message, a routing structure based upon a network configuration and a source network node and destination network node of the first active message;
creating a contention structure created by summing together routing structure elements for active message events;
first calculating, for the first active message, an available bandwidth for use by the message at a path between network nodes utilized by the first active message, the available bandwidth being a function of a level of contention between the first active message and at least a second active message on the path, the level of contention being determined in accordance with the contention structure and the routing structure for the first active message; and
second calculating, for the first active message, based upon the available bandwidth, a modeled communication delay to communicate at least a portion of the first active message. - View Dependent Claims (22, 23, 24, 25, 26, 27, 28, 29, 30)
-
-
31. A computer system including executable program code for generating a delay model in a networked system under a defined workload, the computer system comprising:
-
an input stage that receives a workload description and renders event sequences corresponding to the workload description; and
an evaluation stage that receives the event sequences and renders timing information representing execution of the event sequences in a distributed processing network configuration, the evaluation stage comprising;
a routing structure generator for generating, for a first active message, a routing structure based upon a network configuration and a source network node and destination network node of the first active message;
a contention structure generator for creating a contention structure by summing together routing structure elements for active message events;
a bandwidth availability calculator for first calculating, for the first active message, an available bandwidth for use by the message at a path between network nodes utilized by the first active message, the available bandwidth being a function of a level of contention between the first active message and at least a second active message on the path, the level of contention being determined in accordance with the contention structure and the routing structure for the first active message; and
a delay calculator for second calculating, for the first active message, based upon the available bandwidth, a modeled communication delay to communicate at least a portion of the first active message.
-
Specification