Networking Reference
In-Depth Information
Before diving into details of what drives network performance, we pause to lay the ground work for
some fundamental terminology and concepts. Network performance is characterized by its latency
and bandwidth characteristics as illustrated in Figure 2.2 . The queueing delay, Q(λ) , is a function
of the offered load ( λ ) and described by the latency-bandwidth characteristics of the network. An
approximation of Q(λ) is given by an M/D/1 queue model, Figure 2.2 (a). If we overlay the average
accepted bandwidth observed by each node, assuming benign traffic, we Figure 2.2 (b).
Q(λ) =
When there is very low offered load on the network, the Q(λ) delay is negligible. However, as traffic
intensity increases, and the network approaches saturation, the queueing delay will dominate the
total packet latency.
The performance and cost of the interconnect are driven by a number of design factors,
including topology, routing, flow control, and message efficiency.The topology describes how network
nodes are interconnected and determines the path diversity — the number of distinct paths between
any two nodes. The routing algorithm determines which path a packet will take in such as way as
to load balance the physical links in the network. Network resources (primarily buffers for packet
storage) are managed using a flow control mechanism. In general, flow control happens at the link-
layer and possibly end-to-end. Finally, packets carry a data payload and the packet efficiency determines
the delivered bandwidth to the application.
While recent many-core processors have spurred a 2
increase in the number of
processing cores in each cluster, unless network performance keeps pace, the effects of Amdahl's
Law will become a limitation. The topology, routing, flow control, and message efficiency all have
first-order affects on the system performance, thus we will dive into each of these areas in more
detail in subsequent chapters.
and 4
Layers of abstraction are commonly used in networking to provide fault isolation and device in-
dependence. Figure 2.3 shows the communication stack that is largely representative of the lower
four layers of the OSI networking model. To reduce software overhead and the resulting end-to-
end latency, we want a thin networking stack. Some of the protocol processing that is common
in Internet communication protocols is handled in specialized hardware in the network interface
controller (NIC). For example, the transport layer provides reliable message delivery to applications
and whether the protocol bookkeeping is done in software (e.g., TCP) or hardware (e.g., Infiniband
reliable connection) directly affects the application performance. The network layer provides a logical
namespace for endpoints (and possibly switches) in the system. The network layer handles pack-
ets , and provides the routing information identifying paths through the network among all source,
destination pairs. It is the network layer that asserts routes, either at the source (i.e., source-routed)
Search MirCeyron ::

Custom Search