- Useful Links:
[ RapidIO.org ]
- Objective:
The RapidIOTM Data Streaming Logical layer has been designed to
provide Segmentation and Reassembly (SAR) service to higher layer
applications. According to this service, all application Protocol
Data Units (PDUs) that are greater than the defined MTU size are
first segmented at the sender and then reassembled at the receiver
in a manner that is completely transparent to higher layers. This
simple SAR mechanism, however, has its associated problems. Packet
reassembly requires memory at the receiver. Since a RapidIO
receiver can potentially receive PDUs as big as 64KB from as many
as 64K different senders, the memory requirements for successfully
reassembling packets will be huge (~4GB). Most RapidIO based
hardware chips will have memory that is far less than this minimum
requirement prompting the need to have efficient arbitration
protocols/algorithms that can fairly share memory among competing
traffic streams. The current Data Streaming Logical layer
specification does not have any provision for such arbitration and
solely depends on simple XON/XOFF congestion avoidance mechanisms
(defined in the flow-control logical layer) for this purpose. In
this paper we show that this situation can lead to deadlocks and
system under-utilization. To circumvent this problem, we define a
flow arbitration protocol as an extension to the flow-control
logical layer. This protocol is now a part of the RapidIO 2.0
specification. A defining characteristic of this protocol is that
memory resources can be reserved for Single-PDU transfers as well
as Multi-PDU transfers. Simulation results in Opnet show that this
protocol can increase the system utilization at minimal cost.
- Background:
- Results: [Preliminary]
- Opnet Models:
| Number of sources (N)
| 10
|
| Number of SAR contexts (M)
| 0.5*N = 5
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
- People
- Footnotes
- Sample quotes
- References
- RapidIO technology: Overview and applications.
- RapidIO Interconnect Specification Part 10:
Data Streaming Logical Specification
- RapidIO Interconnect Specification Part 9: Flow Control Logical Layer Extension Specification.
- RapidIO Interconnect Specification Part 1:
IO logical Specication
- RapidIO Interconnect Specification Part 3: Common Transport Specification
- Problems that have research potential:
Below are some problems that I think migh have good research
potential. (In other words, I have not seen any algorithms that can
achieve this).
Is this a simple resource arbitratiion procedure?
Yes, but the situation in which the arbitration occurs is interesting.
1. The receiver does not have any knowledge of the complete packet
size. It only gets packets that are either complete packets or
segments. Thus the receiver has no way to know how many segments are
going to follow the current one. Thus there is no way to decide how
much time the resource (in this case the SAR context) should be
allocated to the sender. This raises some interesting optimization
problems. For example the sender may be very slow and sends very large
number of segments.
In this case, there is no way for the receiver to know
|
