Runtime Networks-on-Chip Performance Monitoring

Networks-on-Chip (NoC) are scalable interconnects for Systems-on-Chip (SoC). An event-based NoC monitoring service has been developed within the DEMONS project which supports runtime observability of NoC behavior. Runtime NoC performance monitoring is an instance of this monitoring service and enables observation of performance measures such as connection latency and link utilization, while a system is actually running. This can for instance be used to obtain network resource information for Quality-of-Service (QoS) management or to support communication centric debugging. This thesis investigates whether runtime NoC performance monitoring is feasible and how it can be used. A framework is presented that identifies NoC performance measures. In order to prove feasibility, runtime NoC performance monitoring is designed and implemented for the Æthereal NoC. For most of the investigated performance measures, generated load heavily depends on the sampling period. Generated load is presented for each of the investigated performance measures. Experiments are conducted for an MPEG application to obtain realistic communication cost figures. These experiments show that the costs for runtime NoC performance monitoring can be low compared to existing traffic, even for very small sampling periods. For instance, link utilization for all links in a 3x1 mesh using a sampling period of 600 ns introduces 1,46 % extra traffic and requires 0,76 % extra energy compared to the communication costs of the MPEG application. This thesis presents NoC congestion control as part of QoS management as an application of runtime NoC performance monitoring. It is well known that an increase in network utilization results in an exponential increase of connection latencies. The goal of NoC congestion control is to bound connection latency by bounding NoC utilization. The suggested control scheme, model predictive control (MPC), combines model based predictions with runtime performance measurements. The control scheme is applied to an Æthereal NoC setup and is implemented to show how the control scheme performs in a realistic environment. For the presented example, latency is reduced from 200 ns per message to 110 ns per message. For the same example we observe a reaction speed of several microseconds when an MPC controller is used that measures link utilization and takes control actions each 600 ns. Experiments with the required link utilization measure have shown 0,19 % of additional traffic and 0,15 % of additional energy for communication, compared to the communication costs of the MPEG application. The main conclusion of this thesis is that runtime NoC performance monitoring is feasible at reasonable costs and that it is usable for, for instance, resource monitoring for QoS management.