CS 591Seminar in Computer Systems

COS 591 - Issues and Discussion Topics

System Area Networking

Routers

The discussion topics of the paper by Dally and Seitz are:

What is wormhole routing
What is virtual channel
Why dimensional routing is deadlock free?
Is dimensional routing deadlock free if there is no virtual channel?
Why do you need virtual channels?

For the rest of the papers, we would like to understand and address the following issues:

How does the router route and how is a route specified
What are the pros and cons of the different approaches to routing
How are virtual channels implemented

arbitration
buffering

How does a system initialize the router?
What do you like and what don't you like in the design?

Finally, we would like to build a list of interesting issues in building a router.

Network Interfaces

The performance requirements for a network interface of a SAN:

Low end-to-end latency (implies low overhead)
High bandwidth (implies small gap)
Support a "convenient" transport protocol layer well

For all papers, we would like to focus on how a particular network interface deal with the following issues:

Communication model
Provide protection among multiple processes
Translation between virtual and physical addresses
Buffer management
Sychronization
Intergration with other components
Operating system support requirements
Initialization

Finally, we would like to build a list of what is good and what is useless in each design.

Transport Protocols and Mechanisms

The focus of the discussions will be on what features to provide at the transport protocol level and how they should be provided:

Protection
Realiable communication
zero-copy and its conditions
data transfer
control transfer
scatter/gather
OS requirements
HW requirements

The performance characteristics includes:

LogP metrics
Latency
Bandwidth

The performance discussion will be focused on tradeoffs with various features.

Architectural Support for Graphics

Graphics Architectural Overview

We plan to focus on the following issues:

How much work is the processing for sort first?
What work does each class of architecture need to do in order to deal with transparency? Recall our class discussion about transparencies vs. z-buffering last time.
Among the 10 hidden-surface algorithms, which one is likely to succeed in the future? If you are not happy with any of them, can you outline what an ideal algorithm looks like?
Among the 10 rasterization techniques, which one is likely to succeed in the future? Do you agree with the performance characterizations? Do you think the characterizations will be changed in the future?
The 10-rasterization paper mentioned that virtual buffer can be used either as a cache or a buffer. What are the pros and cons of the two approaches?

Frame Buffer and Enhanced Memory

The papers in this section cover three areas of topics: the basics of a frame buffer, technology developments in enhanced memory, and advanced features requirements for a frame buffer. Our discussions topics are:

What are CRT's retraces? What does display resolution improves slowly over the years?
In Mary Whitton's paper, she described the pros and cons of using DRAMs and VRAMs to build a frame buffer. The paper was written in 1984, do you think they are still true now? What about in the future?
What are the advantages and disadvantages of integrating logics into DRAMs?
What are the architectural support requirements for texture mapping?
What are the architectural support requirements for supersampling?

Early and Sort-Middle Architectures

The working group posts three general questions:

How much power would you want to put in a display processor and how much in the main CPU?
How much parallelism is there available in graphics and how can one measure the parallelism?
How much can you exploit parallelism with sort-middle architectures and what is the overhead?

Also, we would like to apply the following question to each individual design paper:

Does each design decision still make sense for future graphics architectures?

Other Sort-Middle Architectures

Most of this set of papers describe various SGI machines. There are several questions:

Graphics architecture designers constantly making design tradeoffs between features and tradeoffs. What do you think which features are worth while and which are not worth while?
How does each machine deal with ordering and transparency?
Which features are likely to be the standard features in the future?

A more general question is:

What is a reasonable method to evaluate whether a feature is worth while or not?

Sort-Last Architectures

The issues to be discussed are:

How many kinds of sort-last architectures are there?
What are their advantages and disadvantages?
How do you compare the sort-last and sort-middle architectures? Which one is more sensitive to the depth complexity of the scene? Which one is more scalable?

Sort-First and Chuck-Rendering Architectures

We will be reading eight papers. Here are a number of topics to be discussed:

How do you compare the sort-first architectures with sort-middle and sort-last?
Should the host CPU perform geometry processing or should a graphics accelerator do the job?
How do you compare the stripe approach with chuck rendering?

Back to COS 591 front page | Schedule and Readings | Links | What's New?