- Test 1: test of premium, assured and best-effort with UDP streams and standard queue configurations in the LAN: this test shows the different treatment to different PHBs when traffic is injected for each PHB. Premium traffic always gets the guaranteed bandwidth, independently of the background load. Assuerd traffic always gets at least a minimum amount of capacity as stated in the router and can grab more bandwidth if available by competing with best-effort traffic.
- Test 2: test of a single TCP premium stream with default bucket size of the premium traffic policer. With a TCP source of premium traffic injecting TCP packs of 1500 bytes the default tocket bucket size of 2200 bytes in the router causes starvation.
- Test 3: test of a single TCP premium stream with variable bucket size of the premium traffic policer: by increasing the token bucket size, the tolerance of the policer to TCP bursts increases. Consequently, TCP throughput increases.
- Test 4:
test of TCP premium traffic throughput for different combinations
of tocket bucket size and premium buffer size (guaranteed rate is
constant).
This test shows that TCP throughput can be affected by premium buffer packet drop. Premium buffer drop appears when token_buck_size > premium_buff_size, i.e. when TCP bursts can exceed the premium buffer capacity.
In addition, the test shows that the number of packets dropped by the policer does not depend on the premium buffer size. - Test 5:Impact of token buffer
size and premium buffer size for different premium committed
access rates.
This test shows that the tocken bucket has to be tuned according to the committed access rate. For a constant token bucket size, the number of packs dropped by the token bucket increases. - Test 6:TCP burst tolerance depends on
the token buket size deployed by the premium policer. We define
A token
bucket size "critical" when for smaller values the performance
of a single TCP stream is considerably lower, but a further increase
in bucket depth does not improve it.
The critical bucket depth depends on the rate at which a premium stream is policed.
In this test we compute the "critical" size for a range of different premium rates. - Test 7;
Repetition of test 6 in the WAN (larger
RTT) and for a variable number of TCP connections (1, 5, 10):
with larger RTT for small token bucket size values TCP performance
is better than in the LAN (less sensitive to the burst tolerance
of the policer).
Overall TCP performance does not depend on the number of TCP connections.
- Example of procedure for software upgrade
- Example of initial configuration
- Configuration of token bucket size
- remarks about buffer space management
| |
|
|
|
|
|
Uni. of Stuttgart | Uni. of Athens (NTUA) | for Nuclear Physics and GARR | Uni. of Utrecht | ||
+30.1.772.1865 | |||||
Attn. M Joop Joosten IT Division CH-1211, Geneva 23 Switzerland | Uni. of Stuttgart Attn.: Robert Stoy Allmandring 30 D-70550 Stuttgart Germany | 157 80 Zographou Athens Greece | Attn.: Tiziana Ferrari viale Berti Pichat 6/2, I-40127 Bologna Italy | Attn.: Cees de Laat Princetonplein 5 NL-3584CC Utrecht The Netherlands | |
- Random Early Detection (RED);
L.Linney
(handouts in pdf format) - Differentiated Services on IBM 221X Routers;
L.Linney
(handouts in pdf format) - IBM Diffserv Implementation Experiments; S.Nadas, L.Li, V.Peris - IBM - June 99 (draft-nadas-diffserv-experience-00.txt - gzipped postscript -)
- The Cost of QoS Support in Edge Devices: An Experimental Study; R.Guerin, L.Li, S.Nadas, P.Pan, V.Peris (gzipped postscript)