Some empirical data for the pro-satellite position...
As I mentioned earlier, at JPL we have our mobile satellite protocol
testbed running, and are using it to measure TCP performance under various
modifications. What follows is an example of the type of data we are
gathering.
Our work is meant to reflect a single user on a single hop, such as might
be the case for a mobile user accessing a proxy server at the satellite
groundstation. In this model of satellite access, the user's bandwidth is
fixed (i.e. we do not model contention for the satellite channel).
Our first tests are with large windows as per 1323. For one-way delays
less than about 150ms, the bandwidth*delay product is less than 64k, so the
results are the same with or without large windows. For one-way delays of
150 and above, the results diverge. The table gives the channel
utilization, defined as the number of information bits transmitted in time
T divided by R*T, where R is the channel rate. For 150ms and above,
results are given with and without large windows (XXX/YYY) where XXX is the
channel utilization with large windows, YYY without. Note the greater than
2 to 1 improvement at GEO delays!
Data Rate: 2 Mbps
Channel: Binary Symmetric
TCP Mods: Big Windows
Method: Netperf 10MB transfer.
MSS: 1460
=======================
| One-Way Delay ---> |
=======================
0 50 100 150 200 250 300
|
B 1e-7 92.78 91.68 76.83 63.9/65.7 54.1/48.2 46.4/38.9 46.7/32.6
E
R 1e-8 95.98 94.82 93.26 90.9/77.8 88.7/58.8 82.7/47.7 81.6/39.9
|
v 1e-9 96.94 95.42 93.49 91.4/78.7 89.3/59.9 86.5/48.1 84.9/40.2
0 96.81 95.32 93.46 91.4/78.9 89.0/59.8 87.1/48.1 85.3/40.3
Note that a channel utilization of 80% corresponds to a sustained
throughput of 1,600 kbps, well above the 512kbps value that is sometimes
quoted as the upper limit for TCP over satellites.
Although the channel utilization is shown in the table, the raw data
reflects the latency and the number of retransmissions, which can be used
to calculate the efficiency in terms of the number of total bits
transmitted per data bit.
We are currently working on :
larger initial window size
sack
scps
and hope to have at least preliminary data in time for the december IETF
meeting.
--keith
-----------------------------------------------------------------------------
Keith Scott [email protected]
Jet Propulsion Laboratory
4800 Oak Grove MS 161-260 (Voice) +1.818.354.9250
Pasadena, CA 91109-8099 (FAX) +1.818.393.4643
-----------------------------------------------------------------------------
This archive was generated by hypermail 2b29 : Mon Feb 14 2000 - 16:14:33 EST