Skip to contentSkip to conversation

: Bug fixes in DCCP support

iperf Gerrit Renker 2 participants 3 messages Feb 11, 2008 – Feb 13, 2008

lore
Share & tools
lore
#0 lore raw

I found that iperf still had several bugs, some of these involving the DCCP support.

The bugs have been fixed and the updated version of iperf has been uploaded to
http://www.erg.abdn.ac.uk/users/gerrit/dccp/apps/#iperf

A tarball with all changes is on
http://www.erg.abdn.ac.uk/users/gerrit/dccp/apps/iperf/zip/iperf-2.0.2_DCCP-patched-CBR-continuous.tar.bz2

(NB - test tree also updated today).

Details of changes:
-------------------
* fixed counting bugs when DCCP was used in packet-oriented mode (now the packet count is correct and works across UDP/DCCP);

* enabled UDP-like reporting of statistics also for DCCP;

* rewrote the algorithm to compute the inter-packet-gap in CBR mode:
- in DCCP it tried to stuff the pipe without accounting for the time

     it spent, thus producing more packets than specified;
   - in tests this was observed to cause overflow;
   - the algorithm now measures the actual inter-packet gap of each packet
     and compares it to the target; adjusting for each send time;

- it is more accurate than the previous implementation, in particular so for DCCP.
* replaced the delay_loop(), which was using a spin-loop of constantly calling gettimeofday(), with nanosleep(). While the spinloop worked reasonably in many cases, it sometimes caused unwanted delays of 50milliseconds or more (process preempted?); it does not block signals like nanosleep; and nanosleep uses the hrtimers interface internally (/usr/src/davem-2.6/Documentation/hrtimers/hrtimers.txt)

* separated setting TCP window size and SO_{RCV,SND}BUF settings;

* completely overhauled the IPv6 support, it is in the order of 359 insertions versus 759 deletions; the -V flag is no longer necessary - iperf can now take any v4/v6 hostname, v4/v6 literal address. This required a minor change for the dual test and the tradeoff test (options -2 and -r): they no longer do a check for the reverse-lookup when the reverse connection comes in, instead the peer address is copied from the connection, which will do in almost any conceivable case (doing a reverse lookup leads to complications since it requires complex case distinctions of v4/v6-mapped-v6/v6 addresses).

* IPv4 multicast support did not work well when the server-bind argument (-B) was used on a host with multiple interfaces. This has now been fixed for both IPv4 and IPv6 multicast using the -j' -J' arguments (for join-interface). For example:
iperf -s -u -B 224.1.2.3 # works if there is only one interface
iperf -s -u -B 224.1.2.3 -j eth1 # wanted to join on that one iperf -s -u -B FF00::c0de -J eth2 # IPv6 multicast, same semantics

Hence if anyone is using iperf for DCCP, using the new version is much encouraged, as it will produce more reliable and accurate results.

Still - when encountering any problems with this version of iperf, please do report these.

#1 lore raw

2008/2/11, Gerrit Renker <[email protected]>:

Quoted 52 lines

I found that iperf still had several bugs, some of these involving the DCCP support.

The bugs have been fixed and the updated version of iperf has been uploaded to
http://www.erg.abdn.ac.uk/users/gerrit/dccp/apps/#iperf

A tarball with all changes is on
http://www.erg.abdn.ac.uk/users/gerrit/dccp/apps/iperf/zip/iperf-2.0.2_DCCP-patched-CBR-continuous.tar.bz2

(NB - test tree also updated today).

Details of changes:
-------------------
* fixed counting bugs when DCCP was used in packet-oriented mode (now the packet count is correct and works across UDP/DCCP);

* enabled UDP-like reporting of statistics also for DCCP;

* rewrote the algorithm to compute the inter-packet-gap in CBR mode:
- in DCCP it tried to stuff the pipe without accounting for the time

     it spent, thus producing more packets than specified;
   - in tests this was observed to cause overflow;
   - the algorithm now measures the actual inter-packet gap of each packet
     and compares it to the target; adjusting for each send time;

- it is more accurate than the previous implementation, in particular so for DCCP.
* replaced the delay_loop(), which was using a spin-loop of constantly calling gettimeofday(), with nanosleep(). While the spinloop worked reasonably in many cases, it sometimes caused unwanted delays of 50milliseconds or more (process preempted?); it does not block signals like nanosleep; and nanosleep uses the hrtimers interface internally (/usr/src/davem-2.6/Documentation/hrtimers/hrtimers.txt)

* separated setting TCP window size and SO_{RCV,SND}BUF settings;

* completely overhauled the IPv6 support, it is in the order of 359 insertions versus 759 deletions; the -V flag is no longer necessary - iperf can now take any v4/v6 hostname, v4/v6 literal address. This required a minor change for the dual test and the tradeoff test (options -2 and -r): they no longer do a check for the reverse-lookup when the reverse connection comes in, instead the peer address is copied from the connection, which will do in almost any conceivable case (doing a reverse lookup leads to complications since it requires complex case distinctions of v4/v6-mapped-v6/v6 addresses).

* IPv4 multicast support did not work well when the server-bind argument (-B) was used on a host with multiple interfaces. This has now been fixed for both IPv4 and IPv6 multicast using the -j' -J' arguments (for join-interface). For example:
iperf -s -u -B 224.1.2.3 # works if there is only one interface
iperf -s -u -B 224.1.2.3 -j eth1 # wanted to join on that one iperf -s -u -B FF00::c0de -J eth2 # IPv6 multicast, same semantics

Quoted 5 lines

Hence if anyone is using iperf for DCCP, using the new version is much encouraged, as it will produce more reliable and accurate results.

Still - when encountering any problems with this version of iperf, please do report these.

Good news Gerrit, but in details of changes you mentioned this:

* enabled UDP-like reporting of statistics also for DCCP;

But according to my tests, I had this for UDP

------------------------------------------------------------
Server listening on UDP port 5001
Receiving 1470 byte datagrams
UDP datagram buffer size: 105 KByte (default)
------------------------------------------------------------

[  3] local 127.0.0.1 port 5001 connected with 127.0.0.1 port 48341
[  3]  0.0- 1.0 sec   129 KBytes  1.06 Mbits/sec   0.007 ms    0/   90 (0%)
[  3]  1.0- 2.0 sec   128 KBytes  1.05 Mbits/sec   0.006 ms    0/   89 (0%)

(...)

and for DCCP

------------------------------------------------------------
Server listening on DCCP port 5001
NOTE: running in bytestream-mode (maximum speed)
DCCP datagram buffer size: 105 KByte (default)
------------------------------------------------------------

[  4] local 127.0.0.1 port 5001 connected with 127.0.0.1 port 45839
[  4]  0.0- 1.0 sec  56.7 MBytes   476 Mbits/sec
[  4]  1.0- 2.0 sec  57.1 MBytes   479 Mbits/sec
[  4]  2.0- 3.0 sec  57.0 MBytes   478 Mbits/sec
[  4]  3.0- 4.0 sec  56.9 MBytes   477 Mbits/sec

What is different for the previous version of your patches?

Thanks,

Leandro.

#2 lore raw
In reply to Gerrit Renker
| 
| Good news Gerrit, but in details of changes you mentioned this:
| 
| * enabled UDP-like reporting of statistics also for DCCP;
| 
| But according to my tests, I had this for UDP
| 
| ------------------------------------------------------------
| Server listening on UDP port 5001
| Receiving 1470 byte datagrams
| UDP datagram buffer size:  105 KByte (default)
| ------------------------------------------------------------
| [  3] local 127.0.0.1 port 5001 connected with 127.0.0.1 port 48341
| [  3]  0.0- 1.0 sec   129 KBytes  1.06 Mbits/sec   0.007 ms    0/   90 (0%)
| [  3]  1.0- 2.0 sec   128 KBytes  1.05 Mbits/sec   0.006 ms    0/   89 (0%)
| (...)
| 
| and for DCCP
| 
| ------------------------------------------------------------
| Server listening on DCCP port 5001
| NOTE: running in bytestream-mode (maximum speed)
| DCCP datagram buffer size:  105 KByte (default)
| ------------------------------------------------------------
| [  4] local 127.0.0.1 port 5001 connected with 127.0.0.1 port 45839
| [  4]  0.0- 1.0 sec  56.7 MBytes   476 Mbits/sec
| [  4]  1.0- 2.0 sec  57.1 MBytes   479 Mbits/sec
| [  4]  2.0- 3.0 sec  57.0 MBytes   478 Mbits/sec
| [  4]  3.0- 4.0 sec  56.9 MBytes   477 Mbits/sec
| 
| 
| What is different for the previous version of your patches?
| 

Note the `NOTE' message: you are using in bytestream mode (see http://www.erg.abdn.ac.uk/users/gerrit/dccp/apps/#iperf )

To use packet-oriented mode, use the `-b' option.

The use is as in UDP: without argument, a default CBR of 1 Mbps is used. Otherwise, the argument is CBR in bps (m=Mbps, k=kbps).
In the above, try e.g. -b450m for 450Mbps.

Make sure you use it on both server and client.