46
Network-on-Chip
The GT service guarantees both latency and throughput over a finite time
interval and also supports uncorrupted, lossless, and ordered data transfer. BE
scheme, however, forwards packets as soon as possible, but no guarantees are
given for latency and throughput in general. Vellanki et al. (2004) proposed a
mesh-based router architecture for supporting QoS by modifying a traditional
VC-based router design. In this scheme, two out of four VCs are reserved for
supporting GT services. For high GT load, those traffics are allowed to transmit through the BE VCs, but not vice versa. Reservation-based schemes for GT
traffic generally leads to degradation of average performance for BE traffic.
Nostrum (Millberg et al. 2004) ensures bandwidth for guaranteed throughput traffic by reserving time slots for its transmission on inter-router links. If
no guaranteed throughput traffic is injected into the network, the time slots
are not utilized. Æthereal (Goossens et al. 2005), another mesh-based NoC,
supports guaranteed throughput traffic by utilizing a centralized scheduler
for allocation of link bandwidth. Andreasson and Kumar (2004) proposed a
scheme where BE traffic may use the reserved path when there is no GT traffic
present.
2.7 NI Module
An NI module is used to interface a core with the interconnection network. By
definition, NI module decouples the computation from communication and
performs a protocol conversion between the IP core and the router to which
the core is connected. The Open Core Protocol (OCP) (OCP 2003) is a widely
used interface standard for simplifying the integration task between the IP
cores and the network fabric. Wrapping of IP cores with OCP interface exhibits
a higher reusability and cost-effective plug-and-play-based system implementation. The NI is generic with respect to the network and performs different
services. There are many works reported in the literature on NI design. This
book has taken the work reported by Singh et al. (2007) as an example. They
showed that the NI architecture can be divided into three parts: (1) generic
core interface (GCI), (2) packet maker (PM), and (3) packet disassembler (PD),
as shown in Figure 2.27. The function of each part has been described below.
1. GCI: It lies between the network and the core-specific wrapper like
OCP. It abstracts the network communication protocol from the corespecific wrapper for heterogeneous system implementation. If a new
core is added to the system, the core-specific wrapper views the NI
as a black box.
2. PM: The core-specific wrapper transmits the message to PM memory. The PM performs the following tasks at the source core and
maintains data integrity: