|
|
 |
1. Introduction
SANET2 network throughput measurement was done from 16 to 31 July 2002 during
full production operation. It means that we did not use any special part
of SANET2 network for testing. Measurement was performed between SANET2
nodes at the Computer Center of Zvolen Technical University and the Computer
Center of the Slovak Technical University in Bratislava.
The distance between Zvolen and Bratislava is 265km.
The main goal of our measurement was:
- To verify real capacity of SANET2 network
- To find out error rate during high volume data transfer
- To check out the limit of used network components
2. How we measured
We used two Sun servers Sun Fire V880 with gigabit Ethernet network
cards. These servers were connected together by three ways. The reason was
to get comparative results.
- Local interconnection through L2 switch - both servers were on the
same IP network
- Local interconnection through L3 switch - the servers were connected
to two different IP networks
- Remote interconnection through L3 switches - the servers were connected
in two cities to two different IP networks
| Model |
Sun Fire
V880 |
| Procesor |
4x Ultra SPARC III, 750Mhz |
| Externá cache |
8MB na procesor |
| RAM |
8GB |
| Disks |
6x 36GB Fibre Channel |
| Network interface |
1x Gigabit Ethernet, 1x Fast Ethernet |
|
| OS |
Solaris
9 |
MTU
|
1500 Bytes |
| TCP buffer |
741376 Bytes |
TCP window
|
741376 Bytes |
| Congestion window |
2097152 Bytes |
| Testing software |
NIKHEF ping
NTTCP
Netbench 1.1 |
|
| Software |
Type |
NIKHEF ping
|
ICMP echo/reply |
| NTTCP |
Raw TCP |
Netbench 1.1
|
RAM to RAM data copy with verified checksum |
|
2.1. Measurement No.1 - Local interconnection on L2
In this case both servers were connected to the same
switch Cisco Catalyst 6509 by two 10m long patch cords. Dates were switched
on L2 level during transfer. Both servers were connected to the same IP
network 194.160.10.0.
The main goal of this measurement was to find out the maximum network
throughput between tested servers by excluding delay due to distance.
Two programs measured network throughput NTTCP for raw TCP transfer and
Netbench 1.1 for RAM-to-RAM data copy.
Results
| Test |
èas |
Výsledok |
NTTCP -32768 -n102400
|
36.35 s |
738.4530 Mbit/s |
| Netbench 100MB 1x Stream |
1.247248 s |
67.91260 MB/s (543.30081 Mbit/s) |
Netbench 1000MB 1x Stream
|
15.00486 s |
66.64506 MB/s (533.16049 Mbit/s) |
| Netbench 4000MB 1x Stream |
60.53926 s |
66.07283 MB/s (528.58264 Mbit/s) |
|
2.2. Measurement No.2 - Local interconnection
on L3
In this case, both servers were connected to the same switch Cisco Catalyst
6509 by two ten meters long patch cords. Data were routed on L3 between
servers. The servers were connected to two different IP network 194.160.10.0
and 194.160.11.0.
The main goal of this measurement was to find out and compare throughput
if data are moved on L3 level without influence of IP delay. Delay was measured
in both cases (local L2 and L3 configuration) by NIKHEF ping with the same
results.
Results
| Test |
èas |
Výsledok |
NTTCP -32768 -n102400
|
41.96 s |
639,6412 Mbit/s |
| Netbench 100MB 1x Stream |
1.68066 s |
59.50049 MB/s (476.00394 Mbit/s) |
Netbench 1000MB 1x Stream
|
16.48132 s |
60.67476 MB/s (485.39812 Mbit/s) |
| Netbench 4000MB 1x Stream |
66.99537 s |
59.70562 MB/s (477.64493 Mbit/s) |
|
2.3. Measurement No.3 - Inter city L3 interconnection
In this case servers were connected to two different Cisco Catalyst 6509
switches by two 2m long patch cords. The servers were 265 km away as is
shown on picture above. All network switches were connected through the
single mode optical fiber and long haul GBIC interfaces.
This network path was configured as one VLAN but servers were connected
to two different IP network 194.160.10.0 and 194.160.168.0 and data were
routed on L3 level.
The main goal of this measurement was:
- To verify real capacity of SANET2 network between TU Zvolen and STU
Bratislava
- To check out the influence of 3ms delay caused by 265km distance
on data transfer
- To check out error rate during high volume data transfer
All tests and measurements were performed during working days when average
network load measured by Cisco IOS's was 8%.
Results
| Test |
èas |
Výsledok |
NTTCP -32768 -n102400
|
42.41 s |
632.9231 Mbit/s |
| Netbench 100MB 1x Stream |
1.67023 s |
59.87217 MB/s (478.97734 Mbit/s) |
Netbench 1000MB 1x Stream
|
16.80112 s |
59.51983 MB/s (476.15860 Mbit/s) |
| Netbench 4000MB 1x Stream |
67.16795 s |
59.55222 MB/s (476.41773 Mbit/s) |
| |
| Netbench 100MB 4x Stream |
4.19996 s |
111.99131 MB/s (895.9305 Mbit/s) |
Netbench 1000MB 4x Stream
|
41.37842 s |
105.47521 MB/s (843,8017 Mbit/s) |
|
Error rate
The error rate measured by IOS on Cisco Catalyst 6509 was 0%.
Comparative results
| Test |
L2 lokálne |
L3 lokálne |
L3 vzdialene |
NTTCP -32768 -n102400
|
738.4530 Mbit/s |
639.6412 Mbit/s |
632.9231 Mbit/s |
| Netbench 100MB 1x Stream |
543.30081 Mbit/s |
476.00394 Mbit/s |
478.97734 Mbit/s |
Netbench 1000MB 1x Stream
|
533.16049 Mbit/s |
485.39812 Mbit/s |
476.15860 Mbit/s |
| Netbench 4000MB 1x Stream |
528.58264 Mbit/s |
477.64493 Mbit/s |
476.41773 Mbit/s |
|
3. Summary
We found out the following results:
- If we use bigger TCP window (741376 vs. 49152 default in Solaris9)
the delay caused by distance is negligible.
- During high volume data transfer when average network load is above
80% the error rate is still zero.
- Our network technology - switched gigabit Ethernet is suitable also
for long distance. Both measurements No.2 and No.3 gave us the same
results regardless on distance.
- Network architecture where all backbone routers are connected to
one VLAN allows us to transfer data on L2 level. This concept increases
capacity by 10% because of excluding routing table look up.
- According to our results SANET2 network is able to provide with real
end-to-end network capacity up to 895Mbps in real operation to its users.
Our results are used for first compendium about SANET2 network status.
They proved our network architecture concept and advantages of switched
gigabit Ethernet technology. We recommend the performing of more detail
measurement focused on limits of many others operational systems e.g.
FreeBSD, Linux and Windows. It would be also interesting to perform these
measurements with servers from different vendors and compare results.
To sum it up we can say that the bottleneck of high volume data transfer
is not caused by networking part but by the operational system of end
servers.
4. Measurement team
All tests were performed by:
Sun Fire V880 servers were provided by SUN Microsystems Slovakia and both
technical and system support were provided free of charge by SOMI Systems
a.s. - SUN Enterprise Partner.
|
Version
in Microsoft Word
