Transcript GigaSunet - Uppsala University

GRID Connections in
Börje Josefsson
<[email protected]>
GigaSunet: Nationwide
10 Gbit/s network
• Core network (22 cities):
• 5.270 km of lambdas.
• 27 inter-city segments with
10 Gbit/sec.
• 49 routers (Cisco 124xx).
• Access network (2,5 Gbit/s):
• 25 dark fiber rings.
• 3 segments of 2,5 Gbit/sec.
lambda.
• 1 STM-1 connection.
• 70 routers (Cisco 10720).
Research networks
in Europe
Suggested access scenario
Uplink 2,5 Gbit/sec (redundant)
8 * Fast Ethernet
4 * Gigabit Ethernet
For future
ideas…
Production network
(connected to
University LAN).
GRID
Local peering
IPv6-network
(via UTI)
National research
network (via UTI)
GRID-connection at the edge
• Fiber directly to SUNETs access router.
–
–
–
–
–
There are two access routers at each university.
1 Gbit/sec. [Gigabit Ethernet].
IP version 4 supported today at this speed.
Avoid the local LAN-equipment if possible!
As ”straight” to Your hosts as possible !
• Avoid (cheap) switches etc.!
But it is still slow?!?
For a single session, GigaSunet can be felt1/ as slow!
Speed of light and buffer sizes limits the speed.
Theoretical maximum speed with 32 Kbytes buffer.
Path
Luleå - Umeå
Luleå - Lund
Luleå - Geneva
Luleå - Boston
Luleå - San Diego
RTT
3 ms
32 ms
70 ms
130 ms
205 ms
Max speed
87,3 Mbit/sec
8,2 Mbit/sec
3,7 Mbit/sec
2,0 Mbit/sec
1,3 Mbit/sec
Bwbest = buffer / RTT
1/
Unless You have tweaked Your computer and LAN.
”Fast-and-far”
• You have to think. Twice…
• Tweak Your OS parameters.
• Sweden is approx. 18 ms ”long”.
– 1 Gbit/sec * 36 ms RTT  4,5 megabyte buffer in
the end hosts (usually it is 32 kilobyte…).
• This has nothing to do with the network itself,
this is just how computer communication works.
– It doesn’t help to have a “private” network!
Test results
Luleå-Stockholm (approx. 1.000 km).
Hosts connected to Gigabit Ethernet.
dino# ttcp -s -t -f m -l 61440 -n 20345 wilma.sunet.se
ttcp-t: buflen=61440, nbuf=20345, align=16384/0, port=5001 tcp -> wilma.sunet.se
ttcp-t: socket
ttcp-t: connect
ttcp-t: 1249996800 bytes in 10.35 real seconds = 966181102 bits/sec +++
ttcp-t: 20345 I/O calls, msec/call = 0.52, calls/sec = 1966.33
ttcp-t: 0.0user 1.6sys 0:10real 16% 0i+0d 0maxrss 0+341060pf 41346+4csw
966,2 Mbit/sec.
This is ”real” traffic, not counting the framing and packet overhead.
Can You spot the difference?
Alt. B
Alt. A
Core
switch
Core
switch
Host
Access
router
Data center
switch
Host
Data center
switch
Access
router
Answer:
a) 0,001% more cable in ”B” (compared to total distance) .
b) More (expensive) equipment in ”B”.
c) 75-80% less performance in ”B”.
Conclusion
• SUNET and GigaSunet are ready to handle
this. Now.
• The ”problems” are local at each university.
• We can help You with tuning guidelines and
act as testing partners.
– SUNET test hosts are connected directly to the
core network, without LAN equipment in between.
That helps us isolate “local” problems.
What You should do next..
• Contact Your local network group.
– Get a fiber connection from Your site to one of
the SUNET access routers.
– Have them1 configure the router for Your
connection.
• Connect the GRID equipment, as ”clean” as
possible.
• Tune Your system!
1
/ Done locally at each university, except Uppsala.
Questions?
Börje Josefsson
<[email protected]>