SLIDE 1
1
Status Report -- Gigabit Kits Oakland University
Gigabit Netw ork Technology Workshop July 10-11, 2000
Current Projects
T Improving Performance of Classical IP
between APIC adapter cards
T High Speed DMA Tests with the APIC
adapter cards
T Distributed Processing Using PVM and
MPI
SLIDE 2 2
Can Performance of Classical IP
T Experimental date presented at last
workshop showed:
S maximum data rate surprisingly low S data rate sensitive to record length.
T Experiments used the default send and
receive buffer sizes.
T Can larger socket buffer sizes improve
performance?
Hardw are Configuration
netlab3 WUGS-20
1.2 Gbps ATM 450 MHz
netlab2 netlab4
450 MHz 450 MHz 1.2 Gbps ATM 1.2 Gbps ATM
NetBSD NetBSD Linux Classical TCP/IP
pathclient pathserver
SLIDE 3
3
Description of Experiment
T Send and receive buffer lengths vary over
the set { 65535 49142 32768 16384 8192 4096} (Send and receive buffers are the same length)
T Packet data lengths vary over the same
set.
T 100 packets are sent at each of the 36
length pairs.
Sequence of Events
T Server is started and waits for requests. T Client is started and sends request parameters
to the server.
T Server spawns a child to interact with client
through a TCP/IP session.
T Session ends with statistical report from client,
both human and machine readable.
T Client and server child both terminate. T Server parent waits for another client.
SLIDE 4
4
Sequence of Events
Server Client Server clone
Fastest Data Rates (rw action -- request/response)
Socket buffer length Packet length Data Rate bytes bytes Mb/s
65535 65535 300.064 49142 65535 287.331 49142 49142 266.639 32768 32768 231.849 49142 32768 231.786 65535 32768 231.785 65535 49142 227.832 32768 49142 169.122
SLIDE 5 5
Slow est Data Rates
(rw action -- request/response)
Socket buffer length Packet length Data Rate bytes bytes Mb/s
32768 65535 2.654 16384 65535 2.171 16384 32768 1.171 16384 49142 0.964 49142 16384 0.666 32768 16384 0.665 65535 16384 0.665 16384 16384 0.656
Fastest Data Rates (r action -- unicast)
Socket buffer length Packet length Data Rate bytes bytes Mb/s
65535 65535 300.401 49142 65535 297.544 32768 32768 234.905 65535 32768 234.243 49142 32768 233.558 65535 49142 206.647 8192 8192 159.987 32768 49142 155.779
SLIDE 6
6
Slow est Data Rates (r action -- unicast)
Socket buffer length Packet length Data Rate bytes bytes Mb/s
32768 65535 2.654 16384 65535 2.171 16384 32768 1.171 16384 49142 0.964 49142 16384 0.666 32768 16384 0.665 65535 16384 0.665 16384 16384 0.656
Classical IP Conclusions
T Maximum expected data rate is 300Mb/s. T Actual data rate is sensitive to receive and
transmit socket buffer sizes and packet sizes.
T Most applications conform to the TCP API. T Using TCP data transfer is reliable.
SLIDE 7 7
High Speed DMA tests w ith the APIC
T CSE 647 (Advanced Computer
Communications) term project
T Muthusamy Sivanantham, Rajeshree
Vador, Mahitha Balasubramaniam, Ramya Raghavachar
T Worked with Berkley Shands’ API
Hardw are Configuration
netlab3 WUGS-20
1.2 Gbps ATM 450 MHz
netlab2 netlab4
450 MHz 450 MHz 1.2 Gbps ATM 1.2 Gbps ATM
NetBSD NetBSD Linux VCI 211
DemoLibraryAAL5Read.debug DemoLibraryEndlessAAL5Read.debug
SLIDE 8
8
Performance
(pacer clock at 256 cycles)
T Data rate consistently at 94MB/s
(752Mb/s)
T Small number of samples at 81MB/s
(648Mb/s)
T Very small number of samples at 2MB/s
(16Mb/s)
T 409 buffers -- 36864 bytes
Performance
(Pacer clock at 8192 cycles)
T Data rate consistently at 28.5 MB/s (228
Mb/s)
T Very few samples at 2.6 MB/s (20.8 Mb/s) T 409 buffers -- 36864 bytes
SLIDE 9 9
Pacer clock sets CBR
256 cycles 8192 cycles 28.5 MB/s 94.3 MB/s
APIC API Conclusions
T APIC DMA yields consistently high
performance compared to classical IP over
- ATM. (752Mb/s compared to 300Mb/s.)
T No legacy applications conform to the
APIC DMA API.
SLIDE 10 10
Next Step and status
T Modify PVM and MPI to use the APIC API
for APIC to Process DMA transfers.
T We have just begun to work on these
modifications.
Another Alternative
T Use native ATM API T Create virtual circuit with -paced < rate>
- ption rather than -besteffort
T Native ATM API is close to UDP/IP. T Paced may yield more reliable throughput
than besteffort.
T Performance should still lag compared to
the APIC DMA API.
SLIDE 11
11
Other Projects
T Load Visualizer for the ATM Switch T WUGS Performance Monitor T A Graphical User Interface for Jammer T Gather Statistics during near-overloading
and overloading.
Load Visualizer for the ATM Sw itch
T CSE 647 (Advanced Computer
Communications) term project
T Vamsi Atluri and Naveen Nagaraja T A Java program which plots cell traffic for
selected active virtual circuits in real time.
T Interfaces with Greg Buchman’s program.
SLIDE 12 12
WUGS Performance Monitor
T CSE 647 (Advanced Computer
Communications) term project.
T Ahmad Milhim, Ulvi Bucak, Caijian Pan,
Tom Brusca
T A Tcl/Tk and Perl program which produces
barcharts showing virtual circuit activity in real time.
T Interfaces with Greg Buchman’s program.
A Graphical User Interface for Jammer
T CSE 647 (Advanced Computer
Communications) term project
T Kennet DeMonn, Vilasita Malpeddi, Sridevi
Thamma, Sudha Bhogaraju
T A Java program which presents a GUI for
- Jammer. Jammer is executed each time
via a system() like call.
SLIDE 13 13
Gather Statistics during near-
- verloading and overloading.
T CSE 647 (Advanced Computer
Communications) term project.
T Yangsi Boppana, Mamta Maddireddy,
Niloufer Mathew, Bob Person
T Drove switch into near-overload condition
and measured affect on cross-traffic.
Current and Future Activities
T Distributed computing using APIC DMA
API with PVM and MPI.
T Tunneling ATM over Internet2 using a
tunnel based on udptunnel.
T Scheduling and policing multimedia
multicast and near-multicast streams.
SLIDE 14 14
Distributed Processing
APIC APIC APIC APIC APIC APIC APIC APIC
PC PC PC PC PC PC PC PC I N T E R N E T 2
ATM Tunnel ATM Tunnel
