{"title":"Fast Buffer Memory with Deterministic Packet Departures","authors":"Mayank Kabra, Siddharth Saha, Bill Lin","doi":"10.1109/HOTI.2006.13","DOIUrl":null,"url":null,"abstract":"High-performance routers need to store temporarily a large number of packets in response to congestion. DRAM is typically used to implement the needed packet buffers, but DRAM devices are too slow to match the bandwidth requirements. To bridge the bandwidth gap, a number of hybrid SRAM/DRAM packet buffer architectures have been proposed (S. Iyer and N. Mckeown, 2002) (S. Kumar et al., 2005). These packet buffer architectures assume a very general model where the buffer consists of many logically separated FIFO queues that may be accessed in random order. For example, virtual output queues (VOQs) are used in crossbar routers, where each VOQ corresponds to a logical queue corresponding to a particular output. Depending on the scheduling algorithm used, the access pattern to these logical queues may indeed be at random. However, for a number of router architectures, this worst-case random access assumption is unnecessary since packet departure times are deterministic. One architecture is the switch-memory-switch router architecture (A. Prakash et al., 2002) (S. Iyer et al., 2002) that efficiently mimics an output queueing switch. Another architecture is the load-balanced router architecture (C.S. Chang et al., 2002) (I. Keslassy et al., 2003) that has interesting scalability properties. In these architectures, for best-effort routing, the departure times of packets can be deterministically calculated before inserting packets into packet buffers. In this paper, we describe a novel packet buffer architecture based on interleaved memories that takes advantage of the known packet departure times to achieve simplicity and determinism. The number of interleaved DRAM banks required to implement the proposed packet buffer architecture is independent of the number of logical queues, yet the proposed architecture can achieve the performance of an SRAM implementation","PeriodicalId":288349,"journal":{"name":"14th IEEE Symposium on High-Performance Interconnects (HOTI'06)","volume":"88 8","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2006-08-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"14","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"14th IEEE Symposium on High-Performance Interconnects (HOTI'06)","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/HOTI.2006.13","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 14
Abstract
High-performance routers need to store temporarily a large number of packets in response to congestion. DRAM is typically used to implement the needed packet buffers, but DRAM devices are too slow to match the bandwidth requirements. To bridge the bandwidth gap, a number of hybrid SRAM/DRAM packet buffer architectures have been proposed (S. Iyer and N. Mckeown, 2002) (S. Kumar et al., 2005). These packet buffer architectures assume a very general model where the buffer consists of many logically separated FIFO queues that may be accessed in random order. For example, virtual output queues (VOQs) are used in crossbar routers, where each VOQ corresponds to a logical queue corresponding to a particular output. Depending on the scheduling algorithm used, the access pattern to these logical queues may indeed be at random. However, for a number of router architectures, this worst-case random access assumption is unnecessary since packet departure times are deterministic. One architecture is the switch-memory-switch router architecture (A. Prakash et al., 2002) (S. Iyer et al., 2002) that efficiently mimics an output queueing switch. Another architecture is the load-balanced router architecture (C.S. Chang et al., 2002) (I. Keslassy et al., 2003) that has interesting scalability properties. In these architectures, for best-effort routing, the departure times of packets can be deterministically calculated before inserting packets into packet buffers. In this paper, we describe a novel packet buffer architecture based on interleaved memories that takes advantage of the known packet departure times to achieve simplicity and determinism. The number of interleaved DRAM banks required to implement the proposed packet buffer architecture is independent of the number of logical queues, yet the proposed architecture can achieve the performance of an SRAM implementation