d_l
Barsoom
Premium,MVM
join:2002-12-08
Reno, NV
| ACK Prioritization: a 4100/4100b vs. 5360 shootout
Previous speed tests of simultaneous upload and downloads on various modems by kruser  , deblin , and myself have shown that some of the recent modem models seem to have ACK prioritization (some may have some other download buffering method) and that it is more effective at higher speed profiles. That is the download speed during uploads is more nearly equal to the normal download speed without the upload for faster speed plans. The various test results can be seen in this graph (all tests had an RWIN of ~6xK):
Download throughputs for various modems
Note that the download speeds for any single test are extremely erratic and the data points represent averages of multiple test runs.
I was curious about what the individual packet speeds during simultaneous downloads and uploads were like for the Elite speed plan and about how the modems implement ACK prioritization. To test the speeds, I started a Wireshark packet capture and then began an FTP download speed test from the OOL site during times of the day when the cross-country packet transit isn't impaired. After allowing the FTP download to stabilize, I began a simultaneous FTP upload to an PWP web site on the AT&T backbone. I did this testing for the Speedstream 4100/4100b modems and a 5360 modem using an RWIN of 63888. Then I repeated the tests using an RWIN of 127776. The throughput graphs for the packet captures are as follows:
4100 modem ~64K RWIN
5360 modem ~64K RWIN
4100b modem ~128K Rwin
5360 modem ~128K RWIN
The interval between the red lines was the simultaneous upload. Note that the duration of the upload period is not significant because different-sized binary files were used during the upload tests.
As you can see, most of the individual packets on the 4100/4100b modems stay at full download speeds on the 64K RWIN and even more stay at full speed when the RWIN was doubled. The 5360, on the other hand, has most of its download packets at slower than normal speeds during the upload; however, the download speeds were significantly improved with the larger RWIN. Some speed tests were run with this RWIN and it seems that there is a 2X to 3X improvement in the 5360's download performance with the larger RWIN. This would mean that instead of having 10-20% of normal throughput when the upload is saturated, the 5360 modem on average (there is still quite a bit of variance from test to test) could get 45-65% of normal speeds with the larger RWIN. In comparison, the 4100/4100b modems have very consistent download thoughputs of 90-100% of normal with the larger RWINs.
The bottom line of all this is don't hang on to using your 5360 modem if you do simultaneous download and uploads. If you insist on using the 5360, consider upping your RWIN to the ~128K range (doubling the RWIN yet again provided no additional benefits). The ~128K RWIN size can be beneficial for 4100/4100b users as well, but the improvement is nowhere near as dramatic as it is for modems without ACK prioritization. YMMV! |
|
deblin
Dark Side of the Moon
Premium,MVM
join:2001-09-01
Middletown, DE
 ·Verizon FIOS
 ·Comcast Workplace
 ·DSL EXTREME
| Great work d_l , thanks!
Bottom line - if you have an older modem, consider asking AT&T for the 4100 or go to best buy and shell out the $75 for the modem, it's worth it. Not only do they typically improve line stats and line stability over older modems, they have this added advantage.
--
Hello...is there anybody in there? |
|
d_l
Barsoom
Premium,MVM
join:2002-12-08
Reno, NV
| I want to re-emphasize the YMMV part. The download test I used involved downloading from a relatively high latency, east coast site. The packets had to travel from NY to CA and then back here to NV! This is the type of situation where larger RWINs shine.
The improvement from using a larger RWIN can be seen in the tighter packet speeds on the part of the download before and after the upload portion. More of the packets are nearly at 650 KB/second speed during normal downloading when the RWIN was at ~128K
The ACK prioritization effects clearly interact with and affect the TCP/IP layer. So a larger RWIN might not work as well for someone in the middle of the country that has shorter routes (lower latency) to a download server |
|
kruser
Premium
join:2002-06-01
Chesterfield, MO
clubs:
·AT&T Southwest
| 
Baseline no upload | 
2wire with 65535 RWIN | 
2wire with 127776 RWIN | 
2210 with 65535 RWIN | 
2210 with 127776 RWIN |
The other tests were run with the same download running along with a large FTP upload running simultaneously. |
|
d_l
Barsoom
Premium,MVM
join:2002-12-08
Reno, NV
|  Thanks for doing those! They clearly show that the 2Wire has minimal ACK prioritization. As a very rough, eyeball guesstimate, it looks like the speeds during upload saturation might be about 160 KB/sec with 65K RWIN which would give about 25% or less of normal throughput (160/640). This agrees with your previous tests plotted on my first graph.
The 2210 doesn't look to have much ACK prioritization either. Maybe its speeds during upload saturation might be 150-180 KB which would be 23-28% of normal throughput. This is a little better than your previous testing on the AT&T speed tests, but then again it is really difficult to make a good estimate of average speeds from those throughput graphs.
This pretty much shows that the 2wire 1800HG and the 2210 suck for simultaneous download and uploads unless you use a larger RWIN and that can cause problems. I'm finding that the AT&T speeds tests won't always run in the evenings with an ~128K RWIN. |
|
d_l
Barsoom
Premium,MVM
join:2002-12-08
Reno, NV
| reply to d_l
A point of clarification: I think all the throughput tests (or at least those not done through 2Wires) were performed with a third-party router making the PPPoE connection (all my tests were done this way). This means that the modem does not have to be making the PPPoE connection to provide ACK prioritization.
Exactly how the modem is able to handle this on the PPPoE-encapsulated packets sent to it from the router remains to be explained. |
|
