Υψηλής απόδοσης ενδο-συνδεόμενα ολοκληρωμένα δίκτυα με τη χρήση αλγόριθμου δρομολόγησης με σφαιρική εξισορρόπηση ροής φορτιού
Date Issued
2015
Author(s)
Advisor
Abstract
According to Moore’s law (1965), the density of transistors in an integrated circuit doubles every eighteen months. With integrated circuits containing an increasing amount of transistors, the number of on-chip digital modules and memory units that can fit onto a chip also increases, and as a consequence the performance of computer architectures has been growing at an exponential rate. This trend has led us to the point today where it is possible to pack tens or even hundreds of digital electronic modules onto a single silicon die, comprising multiple tiles, where every tile contains a processing element, a graphic engine, multi-level cache memory, a router, etc. The array of these multiple tiles in a single integrated circuit has given rise to the multi-core processor paradigm.
With multi-core processors there is increasing demand for communication between its tiles. In such a communication-centric system, dedicated wires and limited-connectivity crossbars are no longer adequate in handling the exponentially increasing communication demands, which come in the form of packetized messages exchanges. Alternatively, Networks-on-Chips (NoCs), miniature-scale counterparts of large-scale off-chip networks found in computer clusters, servers and supercomputers, instead provide fast and efficient communication among the various tiles. Besides their high-throughput capabilities, NoCs offer further desirable advantages, such as those of power consumption management, greater tolerance to faults that arise during the operational life of a multi-core processor, modularity, and scalability.
Packetized message delivery among the various tiles of a multi-core processing chip requires the use of a reliable and efficient routing algorithm. The routing algorithm directly impacts the attainable performance of a NoC, since it affects the network’s latency observed when sending messages, and also the network’s bandwidth and effective throughput. Most NoCs tend to use simplified routing algorithms to route packets among tiles, and although they do ensure that the routed data will eventually arrive at their destinations, however, they do not take into consideration the network’s congestion levels and statuses that are currently present. This behavior, unfortunately, limits the network’s effective throughput, and
subsequently constraints the entire performance of the multi-core processor system. In the worst-case scenario, routing algorithms that are oblivious to the current contention state of a network do not aim toward balancing the traffic workload using alternative topology paths, and thus, the congestion in the network is left unmonitored and unconstrained. Next, semi-adaptive routing algorithms, though, do take into consideration the localized, i.e., in their immediate vicinity, network state and try to adapt their routing decisions to the network’s state, but they lack a global and hence optimal view of the network which consequently limits their performance too.
Taking the above shortcomings into consideration, in this thesis, we propose an adaptive routing algorithm with global load-balancing capabilities that takes into consideration the congestion level and state of the traffic’s data flow across the entire network topology. Our goal is to provide the best guidance in routing the data flow across alternative topology paths so that network congestion is restricted, and the network’s throughput and overall performance are maximized. We propose three different versions of this global congestion-aware load-balancing routing algorithm, each with a different achievable performance level. Our experimental evaluation and results, using a cycle-accurate network simulator, show up to 17.5% improvement in the network’s effective throughput when compared to a non-congestion-adaptive routing algorithm.
With multi-core processors there is increasing demand for communication between its tiles. In such a communication-centric system, dedicated wires and limited-connectivity crossbars are no longer adequate in handling the exponentially increasing communication demands, which come in the form of packetized messages exchanges. Alternatively, Networks-on-Chips (NoCs), miniature-scale counterparts of large-scale off-chip networks found in computer clusters, servers and supercomputers, instead provide fast and efficient communication among the various tiles. Besides their high-throughput capabilities, NoCs offer further desirable advantages, such as those of power consumption management, greater tolerance to faults that arise during the operational life of a multi-core processor, modularity, and scalability.
Packetized message delivery among the various tiles of a multi-core processing chip requires the use of a reliable and efficient routing algorithm. The routing algorithm directly impacts the attainable performance of a NoC, since it affects the network’s latency observed when sending messages, and also the network’s bandwidth and effective throughput. Most NoCs tend to use simplified routing algorithms to route packets among tiles, and although they do ensure that the routed data will eventually arrive at their destinations, however, they do not take into consideration the network’s congestion levels and statuses that are currently present. This behavior, unfortunately, limits the network’s effective throughput, and
subsequently constraints the entire performance of the multi-core processor system. In the worst-case scenario, routing algorithms that are oblivious to the current contention state of a network do not aim toward balancing the traffic workload using alternative topology paths, and thus, the congestion in the network is left unmonitored and unconstrained. Next, semi-adaptive routing algorithms, though, do take into consideration the localized, i.e., in their immediate vicinity, network state and try to adapt their routing decisions to the network’s state, but they lack a global and hence optimal view of the network which consequently limits their performance too.
Taking the above shortcomings into consideration, in this thesis, we propose an adaptive routing algorithm with global load-balancing capabilities that takes into consideration the congestion level and state of the traffic’s data flow across the entire network topology. Our goal is to provide the best guidance in routing the data flow across alternative topology paths so that network congestion is restricted, and the network’s throughput and overall performance are maximized. We propose three different versions of this global congestion-aware load-balancing routing algorithm, each with a different achievable performance level. Our experimental evaluation and results, using a cycle-accurate network simulator, show up to 17.5% improvement in the network’s effective throughput when compared to a non-congestion-adaptive routing algorithm.
File(s)![Thumbnail Image]()
Name
Προδρόμου Μιχάλης - Abstract.pdf
Size
144.36 KB
Format
Adobe PDF
Checksum (MD5)
6bf9c601eb9b1038cfe6326228f2f48b