Networking Reference
In-Depth Information
end-to-end flow control,
reliable message delivery
Transport
Transport
routing, node addressing,
load balancing
Network
Network
link-level flow control,
data-link layer reliable delivery
Data Link
Data Link
Physical
Physical
physical encoding (e.g. 8b10b)
byte and lane alignment,
physical media encoding
Interconnection Network
Figure 2.3: The communication stack.
or along each individual hop (i.e., distributed routing) along the path. The data link layer provides
link-level flow control to manage the receiver's input buffer in units of flits (flow control units). The
lowest level of the protocol stack, the physical media layer, is where data is encoded and driven onto
the medium. The physical encoding must maintain a DC-neutral transmission line and commonly
uses 8b10b or 64b66b encoding to balance the transition density. For example, a 10-bit encoded
value is used to represent 8-bits of data resulting in a 20% physical encoding overhead.
SUMMARY
Interconnection networks are a critical component of modern computer systems. The emergence
of cloud computing, which provides a homogenous cluster using conventional microprocessors and
common Internet communication protocols aimed at providing Internet services (e.g., email, Web
search, collaborative Internet applications, streaming video, and so forth) at large scale. While In-
ternet services themselves may be insensitive to latency, since they operate on human timescales
measured in 100s of milliseconds, the backend applications providing those services may indeed
require large amounts of bandwidth (e.g., indexing the Web) and low latency characteristics. The
programming model for cloud services is built largely around distributed message passing, commonly
implemented around TCP (transport control protocol) as a conduit for making a remote procedure
call (RPC).
Supercomputing applications, on the other hand, are often communication intensive and can
be sensitive to network latency. The programming model may use a combination of shared memory
and message passing (e.g., MPI) with often very fine-grained communication and synchronization
Search MirCeyron ::




Custom Search