Database Reference
In-Depth Information
is work in progress in IETF
*
(Internet Engineering Task Force) on the list of
metrics that should be used for evaluating a transport protocol. However, it
is becoming clear that every transport protocol represents a trade-off point
in the vast multidimensional space of evaluation metrics. There is no clear
answer to the general question:
“What is the best protocol?”
as this answer
depends on at least three important factors: What is the definition of
best
for
the given scenario; what is the application type; and what are the deployment,
network, and trac constraints for the scenario?
Many evaluations and comparisons of transport protocols use the transfer of
large data files over wide area networks as a motivating application scenario.
File transfers need reliable transmission, are not highly sensitive to latency or
throughput variations, are usually long-lived, use disk-to-disk transmission,
and the data is organized as a linear sequence. The important measure for
a file transfer application is the total time needed to transfer a file from the
source to the destination.
One of the most utilized data transfer systems for grid and distributed ap-
plications today is GridFTP.
4
,
62
One approach that can be taken to overcome
some of the limitations of TCP is to use parallel connections. GridFTP
3
and
PSockets
76
are early adopters of this mechanism. On public or shared links,
the improvement in data transfer rate may come at the cost of other streams
reducing their transfer rate, which is not always acceptable. Also, on dedi-
cated connections other protocols provide better performance. Recently, the
Globus GridFTP implementation was modified to accept various data trans-
port plug-ins. One of the currently available plug-ins uses the UDT transport
protocol
16
mentioned previously.
One of the practical issues of high-speed protocols is that for some of them
implementations are only available as kernel patches, and switching TCP vari-
ants in the kernel on a shared resource is in most situations undesirable. An
alternative to kernel patches is the implementation of transport protocols in
the user space, using UDP. This is the approach utilized for implementing
transport protocols such as RBUDP or UDT. The UDT library also provides
a framework for custom user-space implementations of various congestion-
control algorithms.
35
The advantage of user-space protocol implementations
is that they can be used in any application. The disadvantage is that they are
more resource intensive; the CPU utilization on the machines is higher than
that of kernel implementations for the same protocol. The packetization effort
is also becoming more significant. The application or the transport protocol
implementation must process each individual packet as it is sent or received
from the network. If the packet size is small, the number of packets becomes
too large for the application to be able to obtain high transmission rates.
Many applications now require the network to support “jumbo” (or 9,000-
byte) packets in order to obtain high transmission rates and to reduce CPU
utilization.
*
http://tools.ietf.org/html/draft-irtf-tmrg-metrics. Accessed July 16, 2009.
