I'm working a transition of a set of test tools from Ubuntu 22.04 to 24.04 and from DotNet 6 to DotNet 8.
I cloned master today and ran into a small issue in make:
In dscp.c, there are a couple of definitions that weren't happy:
Your conditional definition using IPTOS_DSCP_LE (line 98) errors downstream, because IPTOS_DSCP_NQB and NQB2 don't get assigned. My solution for test was to put both in their own #ifndef block to make the compiler happy...
Running a simple test between a Ubuntu 24.04 node (server) and a 22.04 node (client), using the following commands results in a core dump on the server side. The network fabric is 10 Gbps end-to-end.
Server listening on TCP port 5200 with pid 13307
Read buffer size: 64.0 KByte (Dist bin width=8.00 KByte)
TCP congestion control default cubic
Reflected TOS will be set to 0x24 (dscp=9,ecn=0)
TCP window size: 128 KByte (default)
Running without the -e switch eliminated the core dump
When reversed (i.e., 24.04 as client), the -e switch also generated a core dump
Client connecting to 192.168.0.54, TCP port 5200 with pid 13383 (1/0 flows/load)
Write buffer size: 65536 Byte
TCP congestion control using cubic
TOS set to 0x24 (dscp=9,ecn=0) (Nagle on)
TCP window size: 16.0 KByte (default)
Finally, performance is very low. I'm assuming that there has been some changes in buffer sizing behavior (i.e., receive/transmit buffer and RWIN between the 22.04 and 24.04 versions. I'll look into that issue in a follow-up. The commands above, under 22.04, have let RWIN expand to the default Ubuntu value of ~3.2 MB...
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
As to the low performance, I'll collect some data. My VMs are normally able to saturate bidirectionally.
When I performed initial integration of my test tools a year or so ago, there were funny interactions between (some of) the iPerf switches and Ubuntu's TCP stack/kernel parameters (22.04). For example, if I asked for a large RWIN (to approximate existing Riverbed Steelhead configurations), Ubuntu would refuse and set ridiculously small values. I'm going to have to do the same eval with 24.04. I'll keep you posted.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I'm working a transition of a set of test tools from Ubuntu 22.04 to 24.04 and from DotNet 6 to DotNet 8.
I cloned master today and ran into a small issue in make:
In dscp.c, there are a couple of definitions that weren't happy:
Your conditional definition using IPTOS_DSCP_LE (line 98) errors downstream, because IPTOS_DSCP_NQB and NQB2 don't get assigned. My solution for test was to put both in their own #ifndef block to make the compiler happy...
Running a simple test between a Ubuntu 24.04 node (server) and a 22.04 node (client), using the following commands results in a core dump on the server side. The network fabric is 10 Gbps end-to-end.
Server: iperf -i 1 -s -p 5200 --tos-override 36 -e -l 64K
Client: iperf -S 36 -t 30 -i 1 -p 5200 -c 192.168.1.68 -e -l 64K
Server side output:
Server listening on TCP port 5200 with pid 13307
Read buffer size: 64.0 KByte (Dist bin width=8.00 KByte)
TCP congestion control default cubic
Reflected TOS will be set to 0x24 (dscp=9,ecn=0)
TCP window size: 128 KByte (default)
*** buffer overflow detected ***: terminated
Aborted (core dumped)
Running without the -e switch eliminated the core dump
When reversed (i.e., 24.04 as client), the -e switch also generated a core dump
Client connecting to 192.168.0.54, TCP port 5200 with pid 13383 (1/0 flows/load)
Write buffer size: 65536 Byte
TCP congestion control using cubic
TOS set to 0x24 (dscp=9,ecn=0) (Nagle on)
TCP window size: 16.0 KByte (default)
*** buffer overflow detected ***: terminated
Aborted (core dumped)
Finally, performance is very low. I'm assuming that there has been some changes in buffer sizing behavior (i.e., receive/transmit buffer and RWIN between the 22.04 and 24.04 versions. I'll look into that issue in a follow-up. The commands above, under 22.04, have let RWIN expand to the default Ubuntu value of ~3.2 MB...
Thanks for this. Can you try master now?
1. Code commit to protect against connect report buffer overflows (in master)
2. Code commit for ifndef
I can't speak to low performance. My systems drive 10G links fine. Let me know what you figure out or if I can help diagnose somehow.
Last edit: Robert McMahon 2024-07-27
Ok, I think I found the buffer overflow. It's in the CCA parts. Hopefully fixed now.
I'll take a look later today...
As to the low performance, I'll collect some data. My VMs are normally able to saturate bidirectionally.
When I performed initial integration of my test tools a year or so ago, there were funny interactions between (some of) the iPerf switches and Ubuntu's TCP stack/kernel parameters (22.04). For example, if I asked for a large RWIN (to approximate existing Riverbed Steelhead configurations), Ubuntu would refuse and set ridiculously small values. I'm going to have to do the same eval with 24.04. I'll keep you posted.
Sorry for the late response, but that got it.