Methods, systems, and computer readable media for analyzing data center congestion control
A method for analyzing data center congestion control includes capturing packets traversing network switches in a data center configured with data center quantized congestion notification (DCQCN) parameters to regulate congestion in the data center. The method further includes generating, from the captured packets, a congestion notification packet (CNP) profile for the data center and a visualization of real traffic rates in the data center. The method further includes configuring a DCQCN performance predictor with the DCQCN parameters and providing the CNP profile as input to the DCQCN performance predictor. The method further includes simulating, by the DCQCN performance predictor, congestion control of the data center and generating, as output, a visualization of simulated ideal traffic rates in the data center given the CNP profile and the DCQCN parameters.
1 . A method for analyzing data center congestion control, the method comprising:
capturing packets traversing network switches in a data center configured with data center quantized congestion notification (DCQCN) parameters to regulate congestion in the data center;
generating, from the captured packets, a congestion notification packet (CNP) profile for the data center and a visualization of real traffic rates in the data center;
configuring a DCQCN performance predictor with the DCQCN parameters and providing the CNP profile as input to the DCQCN performance predictor; and
simulating, by the DCQCN performance predictor, congestion control of the data center and generating, as output, a visualization of simulated ideal traffic rates in the data center given the CNP profile and the DCQCN parameters.
2 . The method of claim 1 wherein capturing packets traversing network switches in the data center includes capturing packets relating to an artificial intelligence model training workload in the data center.
3 . The method of claim 1 wherein generating the CNP profile includes extracting information from headers of the packets.
4 . The method of claim 3 wherein the information from the headers of the packets includes frame length, source Internet protocol (IP) address, destination IP address, queue pair ID (QpID), explicit congestion notification (ECN) marking, opcode and packet timestamp.
5 . The method of claim 1 wherein generating the CNP profile includes generating a multiple CNP profile that instructs the DCQCN performance predictor to simulate transmission of multiple CNPs.
6 . The method of claim 1 wherein generating the CNP profile includes generating a single CNP profile that instructs the DCQCN performance predictor simulate transmission of a single CNP.
7 . The method of claim 1 wherein generating the visualization of real traffic rates includes generating a graph of traffic rates versus time in the data center.
8 . The method of claim 1 wherein configuring the DCQCN performance predictor with the DCQCN parameters includes configuring the DCQCN performance predictor with the same DCQCN parameters used to regulate congestion in the data center.
9 . The method of claim 1 wherein simulating the congestion control includes executing a model that predicts a target rate and a current rate given the CNP profile and the DCQCN parameters.
10 . The method of claim 1 wherein generating the visualization of simulated ideal traffic rates includes generating a graph of the simulated ideal traffic rates over time in the data center.
11 . A system for analyzing data center congestion control, the system comprising:
at least one processor and a memory;
a packet capture module executed by the at least one processor for capturing packets traversing network switches in a data center configured with data center quantized congestion notification (DCQCN) parameters to regulate congestion in the data center;
a congestion signal extractor executed by the at least one processor for generating, from the captured packets, a congestion notification packet (CNP) profile for the data center and a visualization of real traffic rates in the data center; and
a DCQCN performance predictor executed by the at least one processor for receiving, as input, the DCQCN parameters and the CNP profile, simulating congestion control of the data center and generating, as output, a visualization of simulated ideal traffic rates in the data center given the CNP profile and the DCQCN parameters.
12 . The system of claim 11 wherein the packets traversing network switches in the data center includes packets relating to an artificial intelligence model training workload in the data center.
13 . The system of claim 11 wherein the CNP profile includes information from headers of the packets, including frame length, source Internet protocol (IP) address, destination IP address, queue pair ID (QpID), explicit congestion notification (ECN) marking, opcode, and packet timestamp.
14 . The system of claim 11 wherein the CNP profile includes a multiple CNP profile that instructs the DCQCN performance predictor to simulate transmission of multiple CNPs.
15 . The system of claim 11 wherein the CNP profile includes a single CNP profile that instructs the DCQCN performance predictor to simulate transmission of a single CNP.
16 . The system of claim 11 wherein the visualization of real traffic rates includes a graph of real traffic rates versus time in the data center.
17 . The system of claim 11 wherein the DCQCN parameters received as input to the DCQCN performance predictor include the same DCQCN parameters used to regulate congestion in the data center.
18 . The system of claim 11 wherein the DCQCN performance predictor is configured to execute a model that predicts a target rate and a current rate given the CNP profile and the DCQCN parameters.
19 . The system of claim 11 wherein the visualization of simulated ideal traffic rates includes a graph of the simulated ideal traffic rates over time in the data center.
20 . A non-transitory computer readable medium having stored thereon executable instructions that when executed by a processor of a computer control the computer to perform steps comprising:
capturing packets traversing network switches in a data center configured with data center quantized congestion notification (DCQCN) parameters to regulate congestion in the data center;
generating, from the captured packets, a congestion notification packet (CNP) profile for the data center and a visualization of real traffic rates in the data center;
configuring a DCQCN performance predictor with the DCQCN parameters and providing the CNP profile as input to the DCQCN performance predictor; and
simulating, by the DCQCN performance predictor, congestion control of the data center and generating, as output, a visualization of simulated ideal traffic rates in the data center given the CNP profile and the DCQCN parameters.