As the optimal radix for switches increases due to the benefits 
in lower latencies, overall reduction in cost and power 
consumption; the traditional switch architectures are no longer 
valid because of either low-performance or non-scalability with 
the number of ports.

This dissertation proposes a new switch architecture suitable for 
high-radix switches called Partitioned Crossbar Input Queued 
(PCIQ) that deals with one of the main constraints in high-radix 
switch design, the excessive memory requirements. Also, in 
general terms, PCIQ forms a new family of switch 
microarchitectures.

PCIQ relies on a smart partition of the crossbar into 
sub-crossbars, thus requiring less memory resources than other 
proposals for high-radix, yet obtaining high-performance and also 
increasing the arbiter efficiency. PCIQ uses two round-robin 
packet-based arbiters (one for each crossbar) that exhibit a 
linear cost and a logarithmic response time as the radix of the 
switch increases.

Here it is shown that PCIQ exhibits a cost (measured in terms of 
memory requirements, crossbar complexity and arbiter complexity) 
similar to or lower than basic organizations like CIOQ. However, 
it is able to achieve maximum switch efficiency for uniform 
traffic distribution, thus leveling costly organizations like BC.

The other big issue on high-radix switches is the HOL blocking 
problem, which reduces dramatically the switch performance. 
Traditional solutions for removing the HOL blocking problem were 
based on VOQ schemes, but having high number of ports on a 
high-radix switch prevents the use of any of them. In this 
dissertation, a new congestion management technique has been 
proposed. This solution is called RECN-IQ, is specific for IQ 
switches and differs from the original RECN idea (suitable only 
for CIOQ switches) in being highly efficient and simple to 
implement, reducing the memory requirements to the maximum. 
RECN-IQ introduced by first time a novel statistical approach for 
detecting congestion using just a single queue per input port.

By combining the PCIQ microarchitecture with RECN-IQ, a new 
switch architecture (called here PCIQ-enhanced) is derived and 
evaluated in this dissertation. The PCIQ switch architecture 
inherits the benefits of the Partitioned Crossbar 
microarchitecture in reducing the memory requirements for 
high-radix designs with the power of a congestion management 
technique that removes the HOL blocking dynamically, thus 
achieving maximum switch performance under all types of traffic.

We have seen that in modern interconnection networks it is 
mandatory the use of an effective congestion management technique 
in order to keep network performance at maximum level under 
congestion situations. Therefore, in this dissertation we 
describe the new congestion management technique (RECN-IQ) 
suitable for any type of IQ switches (which includes PCIQ). The 
idea behind RECN-IQ is, starting with a simple IQ switch with a 
single queue per input port, to add some extra queues dynamically 
allocated for congested packets. Congestion is detected as soon 
as HOL blocking begins to act, setting aside (in those extra 
queues known as SAQs) the congested packets in an efficient 
manner. Therefore, HOL blocking is completely eliminated (as 
proven by the simulation results). The hardware requirements for 
RECN-IQ, as we have seen, are reduced, making feasible its 
implementation on any IQ-based switch architecture like PCIQ.

In order to prove that fact, a feasible and realistic switch 
architecture implementing RECN-IQ has been proposed and described 
in detail. Moreover, we have detailed every functional unit and 
structure required to implement RECN-IQ on an Input-Queued switch 
architecture. This is the first time since the RECN proposal back 
in 2005 that a RECN-like congestion management technique has been 
implemented in such a detailed level.

Results proved that by using RECN-IQ switches, the network will 
benefit from low cost switches and high-efficiency under any type 
of traffic pattern or network circumstances. All this makes the 
network predictable and stable in performance, no more drops in 
throughput because of congestion.