RECN-IQ: A Cost-Effective Input-Queued Switch Architecture with Congestion Management * Gaspar Mora† Pedro J. Garcia‡ Jos´ e Flich† Jos´ e Duato† †Dept. of Computer Engineering ‡ Dept. of Computing Systems Universidad Polit´ ecnica de Valencia, Spain Universidad de Castilla-La Mancha, Spain {gmora, jflich, jduato}@gap.upv.es pgarcia@dsi.uclm.es Abstract As the number of computing and storage nodes keeps in- creasing, the interconnection network is becoming a key element of many computing and communication systems, where the overall performance directly depends on network performance. This performance may dramatically drop during congestion situations. Although congestion may be avoided by overdimensioning the network, the current trend is to reduce overall cost and power consumption by reduc- ing the number of network components. Thus, the network will be prone to congestion, thereby becoming mandatory the use of congestion management techniques. In that sense, the technique known as Regional Explicit Congestion Notification (RECN) completely eliminates the Head-of-Line (HOL) blocking produced by congested pack- ets, turning congestion harmless. However, RECN has been designed for switches with queues at input and output ports (CIOQ switches), thus it can not be directly applied to other types of switches. Additionally, the method RECN uses for detecting congestion requires several detection queues that increase the memory requirements and thus switch cost. Thus, we completely redefine the RECN mechanism in order to achieve different goals. First, we adapt RECN to a switch organization with queues only at input ports (IQ switches). These switches are simpler and cheaper to pro- duce than CIOQ ones. Second, we propose a new method for detecting congestion that does not require several detec- tion queues, thereby reducing RECN memory requirements. These improvements lead to achieve a cost-effective switch organization that derive maximum performance even in the presence of congestion. Also, we present in detail a realistic switch architecture supporting the new mechanism. Results demonstrate that the new RECN version in an IQ switch achieves maximum network performance in all the analyzed situations. These results have been obtained with * This work was supported by Spanish MEC under Grant TIN2006- 15516-C04, by Junta de Comunidades de Castilla-La Mancha under Grant PBC-05-005 and by CONSOLIDER-INGENIO 2010 under Grant CSD2006-00046. a reduction factor of data memory requirements of 5 with respect to the previous RECN mechanism in CIOQ switches. 1. Introduction High-performance interconnection networks are nowa- days present in a wide variety of computing and communi- cation systems: massive parallel processors, local and sys- tem area networks, clusters of PCs and workstations, IP routers, and, recently, inside the chips (Networks on Chip). In such environments, as the number of processing and stor- age nodes increases, the interconnection network plays a prominent role in the performance achieved by the whole system. One of the main concerns interconnect designers faced during the last years has been network congestion. Con- gestion occurs when several flows of packets simultane- ously and persistently request the access to the same net- work resources (typically, a switch output port). In these cases, any packet not granted will block, and will remain stored 1 in a queue until its request is attended. This may cause the appearance of the phenomenon known as Head- Of-Line (HOL) blocking, that occurs when a packet at the head of a queue blocks, preventing the rest of packets in the same queue from advancing, even if they request avail- able resources. When congested packets block and pro- duce HOL blocking to non-congested ones (those packets belonging to flows that do not contribute to congestion), non-congested flows advance at the same speed as con- gested flows, thereby severely degrading network perfor- mance and eventually collapsing the network (the effect is rapidly spread over the entire network). Although the negative consequences of congestion have been always evident, congestion has not been considered a critical problem until recently, due to several reasons. For 1 We assume lossless networks, where blocked packets are never dis- carded. Note that most current interconnects (Myrinet 2000, Quadrics, InfiniBand, Advanced Switching, etc.) are lossless. 2007 International Conference on Parallel Processing (ICPP 2007) 0-7695-2933-X/07 $25.00 © 2007 Authorized licensed use limited to: UNIVERSIDAD POLITECNICA DE VALENCIA. Downloaded on November 4, 2009 at 11:17 from IEEE Xplore. Restrictions apply.