Re: Aurora implementation
- From: Austin Lesea <austin@xxxxxxxxxx>
- Date: Wed, 30 Aug 2006 11:06:24 -0700
vt2001cpe,
It seems that if bit errors are not acceptable, as you say, they WILL
happen (if not because of marginal signal integrity of cables,
connectors, etc. then because of EMI/RFI ... in efect "**it happens").
Given a bit error is unavoidable, then that says you MUST use FEC, or a
acknowledge, negative acknowledge (resend) protocol.
A bit error causing the link to loose frame synchronization, or channel
synchronization is something I have not studied for the Aurora core, and
you could email a webcase to ask them exactly that question: "how does
aurora deal with a bit error that may cause frame, or channel
synchronization? Can that even happen?" It may not be possible to
corrupt a 8b/10b character such that you lose sync...after all, it is a
well tried and well thought out old coding (by IBM), and they were
always good at making channel coding such that loss of link sync was not
possible, even at very high error rates...
For example to lose frame sync on a T1 or E1 link requires an error rate
of ~ 1E-3 or worse. Systems are usually designed such that single
isolated bit errors don't cause problems (except to data).
Austin
vt2001cpe wrote:
Here is my potential setup:.
1. N-boxes connected to a single box (Spoke to Hub config) via RocketIO
running the simplex Aurora protocol.
2. Each channel would consist of four 3.125Gb/s lanes bonded together.
3. Time aligned samples are transmitted from each spoke to the hub.
Here are my assumptions, please correct me if I am wrong:
1. If my data throughput is less than the max throughput of the link,
the Link Layer will add filler data during the idle periods that will
be stripped out on the receiver side.
2. The elastic buffers on the TX and RX sides are filled and emptied,
respectively, by me.
3. If data were continuously fed into the TX buffer, sent, and
correctly received by the receiver, the RX buffer would not reach an
empty state.
4. I do not have to send alignment characters after the channel has
been successfully initialized.
I am concerned that if a sample took a bit hit, which caused the sample
to be incorrectly interpreted as a idle or alignment character, the
sample would basically be thrown away. This results in my data no
longer being time aligned, which is catastrophic for this application.
Any insight or suggestions? Is it possible to pass all data
transmitted, user data and special characters to the buffers?
Given the number of channels, the length of operation, and the rate of
transmission, I believe bit errors will be unavoidable. Incorrect data
is relatively acceptable, but skewed data is not.
Thanks!
- References:
- Aurora implementation
- From: vt2001cpe
- Aurora implementation
- Prev by Date: Re: Xilinx - no secret, you are not to use the PMV primitive
- Next by Date: Re: placing addiional caps across existing caps to reduce noise
- Previous by thread: Aurora implementation
- Next by thread: Re: Aurora implementation
- Index(es):
Relevant Pages
|
