Supporting Multimedia Communication over a Gigabit Ethernet NetworkVARUN PIUS RODRIGUES

Need for Gigabit Networks

Streaming Multimedia

High performance distributed computing

Virtual Reality

Distance learning

Development and Design Challenges

Hardware Components:

Gigabit Network Interface Card

A buffered hub

Gigabit routing switch

File Server within LAN: Building pilot workgroups within LAN

Integration of gigabit switches

Design Challenges:

Simplicity

End-to-end solutions

Extended to include emerging multimedia applications

Design Features of Gigabit NIC

802.3 Frame compatibility

Design challenge of the MAC ASIC to operate the GNIC

Reducing host CPU utilization through Descriptor-based DMA

IEEE 802.3z Standard

Standard for MAC and PHY layers

PHY Layer:

Fiber: 1000 BASE-SX (multi-mode) and 1000 BASE-LX (single-mode)

Copper-based: 1000 BASE-CX (twinax cable)

MAC Layer:

Identical to one defined for 10 Mbps and 100Mbps Ethernet

Fields: DA, SA, LEN, DATA, FCS

Designing GNIC

Architecture Consideration: 64-bit 66 MHz GNIC-II

Theoretical bandwidth: 4 Gbps vs 1 Gbps

Practical bandwidth: 3 Gbps vs 800 Mbps

Consists of:

Application-specific Integrated Circuit (ASIC)

Packet buffer memory

Serializer/Deserializer chip

Physical layer components

ASIC

Consists of: PCI

Pair of DMA for controllers: for Rx and Tx

Pair of FIFO connected to external FIFO interface

GNIC

Reducing CPU utilization:

Accesses host memory directly through Descriptor-based DMA

Transfer Chaining: transferring arbitrary nos of packets from host memory to GNIC

Adaptability of interrupt rate of host to network load

Design Features of Buffered Gigabit Hub

Full-duplex: for eliminating CSMA/CD collision

Congestion Control: to avoid frame dropping

Round-Robin scheduling: to prevent “packet clumping”

Performance issues with CSMA/CD

Performance highly dependent on ratio of propagation delay to average packet transmission time

Two ways to solve it:

Increase minimum packet length

Decrease length

Length increased through virtual collisions

Full Duplex Repeater

Achieves switching and shared design concepts for switch-like performance while maintaining the cost of shared hub

Provides maximum throughput, collisionless forwarding, and congestion control

Logical flow of frames: Input -> Forwarding path -> Output

Input: Passes though buffer PHY and MAC before being queued in buffer; congestion control informs end system to slow down incase capacity of buffer is about to be reached

Forwarding path: Implements round robin scheduling to determine which port will send data

Output: includes a buffer, MAC and PHY with congestion notification

Design Features of Gigabit Routing Switch

Architecture issues

Parallel access shared memory architecture

Priority Queue Design

Architecture of Gigabit Routing Switch PE-4884

Consists of 12 channel cards, 1 EMM card for the chassis to function and 1 EMM card management, policy and routing table redundancy

Channel cards: Supports connectivity to Ethernet, fast Ethernet, FDDI etc Sending and receiving data through physical interfaces and system packet

memory

Performing routing and switching address lookups

Enforcing layer 4 policies; collecting management statistics, etc

Every channel card connected to central memory through 2 full-duplex gigabit channels

Contd…

Architecture of Gigabit Routing Switch PE-4884

Memory Architecture:

Uses parallel memory architecture for high speed performance

Limitations of cross-bar architecture:

Port-based memory

Head-of-line blocking

Difficult to provide QoS support

Shared memory bus architecture overcomes these limitations

Packet flow:

Address Resolution Logic ASIC on channel card evaluates the destination address

ARL signals memory control cards that a packet must be sent to which port

Frame Management

Supporting Distributed Multimedia Applications

Challenges in On-Demand Video:

Large data size

Real-time constraint

Supporting concurrent accesses

Consideration for connection setup:

Implementation of RSVP scheme

New emerging standards such as IEEE 802.1p and 802.1Q

Integrated solutions with policy-based QoS

Experimental Results

Two major goals:

To test how a gigabit LAN performed in the basic metrics of throughput

To examine how concurrent video delivery is supported over gigabit LAN

Experiments designed to determine bottlenecks in current system and identify bounds on performance

End-to-end performance is evaluated

Experimental Setup

Hardware setup:

Pentium II 233 Mhz and Pentium Pro 300 Mhz

Running either Linux or NT

Benchmarking utilities: netperf (for max throughput) and netbench (for avg throughput)

Netperf:

Consists of 2 process: netserver and netperf

Netperf connects to remote system running an instance of netserver and uses control connection to send parameters

TCP connection using BSD sockets

Separate connection for measurement

Contd..

Experimental Setup

Netbench:

Measures how well a file server handles I/O requests

Each client tallies how many data moves to and from server

Results

Maximum throughput results:

Message packet size varied from 2048 bytes to 4 Mbytes.

After 16KByte packet size is reached, there is slight drop in performance and then again increase to give peak throughput of 190Mbps

Performance peak in Linux system; NT system peak throughput around 90Mbps

To investigate, NT experiment performed on different machine; peak throughput of 180 Mbps attained

SUN’s Solaris attained peak performance of 488 Mbps

Raw device testing attained performance of 700-800 Mbps at hardware level

Average throughput results:

Maximum server throughput of 157 Mbps with 3 clients

High Quality Streaming Videos

Criteria for experiment to measure concurrent access for on-demand video:

Buffering Scheme:

Used 2 buffer scheme at server: one to retrieve video frames from server system and other to transmit it to client

Client also uses 2 buffer scheme: one for the network and other to display the frame

Performance metrics for the measurement:

Need to determine maximum number of concurrent accesses that can be supported by the network

Need to calculate jitters (number of miss deadline retrievals)

Results: On-Demand Video Streaming

Range specifications of MPEG-2 between 4 and 32 Mbps were emulated

Number of active processes were increased until system could no longer provide acceptable QoS (jitter less than 1%)

Due to bus contention, 128 Mbps achieved for 32 streams of 4 Mbps videos

Buffer size is another performance bottleneck in addition to network throughput

Questions?

You may send me a mail on my UF mail in case of any issues you need to communicate with me

UF mail: varunpius@ufl.edu