As I near the end of my three month window for writing SPDY Book, I am beginning to fret over the things that I have not covered. The biggest, most worrisome outstanding bit for me at this point is lack of counterpoint.
So up tonight, I am ready to give SCTP (Stream Control Transmission Protocol) a whirl. SCTP is a transport layer protocol, so it is not a direct competitor to SPDY, which sits atop HTTP in the application layer). It still aims to solve some of the same problems as SPDY so it is worth a go.
As a potential replacement for TCP/IP, my initial impressions are that it stands no chance. It takes me quite a bit of digging before I even come across a simple example of how to test this stuff. How can it possibly a serious possibility of being installed on computers across the world?
First up, I install
lksctp-tools
. Fortunately, no compilation is required on Ubuntu/Debian. Rather I can pull it directly from apt-get:sudo apt-get install lksctp-toolsBefore actually trying things out, I check that Wireshark can grok SCTP. Under Analyze, I choose Enabled Protocols and see:
Cool beans. Wireshark rocks.
I also check through the filtering options on Wireshark:
I do not think that I will need them tonight. Good to know they are there. For now, I am content to capture anything SCPT, so a Wireshark capture filter of
sctp
will suffice.With a live SCTP capture running in Wireshark, I start up my server. I use the
sctp_darn
utility from lksctp-tools:➜ spdybook git:(master) sudo sctp_darn -H 0 -P 250 -lIn a separate window, I start my client (the args are directly from the linked tutorial):
sctp_darn listening...
➜ spdybook git:(master) sudo sctp_darn -H 0 -P 260 -h 127.0.0.1 -p 250 -sChecking Wireshark, I see that no traffic has been sent:
sctp_darn ready to send...
0:260-127.0.0.1:250>
I am unsure if this is a feature of SCTP or just how the
sctp_darn
utility works. Normal TCP/IP sockets would have sent data at this point. Regardless, I move ahead with my investigation.Back in my client window, I send data:
0:260-127.0.0.1:250> Hello SCTP!!!In the server terminal, I find:
Recieved SCTP_COMM_UP
New connection, peer addresses
127.0.0.1:250
192.168.0.20:250
0:260-127.0.0.1:250>
Recieved SCTP_COMM_UPWell that wasn't so bad after all!
NOTIFICATION: ASSOC_CHANGE - COMM_UP
SNDRCV
sinfo_stream 0
sinfo_ssn 0
sinfo_flags 0x0
sinfo_ppid 0
sinfo_context 0
sinfo_tsn 238345625
sinfo_cumtsn 0
sinfo_assoc_id 4
DATA(15): Hello SCTP!!!.
Checking things out in Wireshark, I see the famed 4 way, SCTP handshake (avoids TCP's SYN flooding):
Last up tonight, I give this a try over the network. On my server, I run the same server command as above. Similarly, on my laptop, I run the same command as early, only this time I supply my server's hostname:
And, amazingly, it just works on my server:
➜ spdybook git:(master) ✗ sudo sctp_darn -H 0 -P 260 -h londo -p 250 -s
[sudo] password for cstrom:
sctp_darn ready to send...
0:260-londo:250> Hello!
Recieved SCTP_COMM_UP
cstrom@londo:~$ sudo sctp_darn -H 0 -P 250 -lI do not know why I expected trouble there, but it does just work on my Ubuntu machines. I have to admit that there may be something to this SCTP stuff after all.
sctp_darn listening...
Recieved SCTP_COMM_UP
NOTIFICATION: ASSOC_CHANGE - COMM_UP
SNDRCV
sinfo_stream 0
sinfo_ssn 0
sinfo_flags 0x0
sinfo_ppid 0
sinfo_context 0
sinfo_tsn 184442484
sinfo_cumtsn 0
sinfo_assoc_id 1
DATA(32): Hello Londo, I'm talking SCTP!.
That will do for tonight. Up tomorrow, I will see if I can put SCTP in situations more akin to what SPDY is trying to solve.
Day #69
If I can cast a vote for you, I argue that SCTP is a waste of time. It's a kernel level change which means most people can't use it. If most people can't use it, nobody will optimize it. If nobody will optimize it, it will only be so-so.
ReplyDeleteBut even worse than being a kernel-level change, its also bound to IPv6! Why? IPv4 today depends on NAT (who doesn't use NAT?) And NAT *only* works with TCP and UDP.
So, if you're planning on doing SCTP research, I argue it is a total waste of time. If you want to research SCTP over UDP, then that might be more interesting.
Contact me offline if you want me to spill my views on all this stuff. But, after having dabbled with SPDY, SCTP, TCP, and even custom UDP transports, I'm pretty sure I'm right :-)
Heh. I'm not sure which Mike you are -- your Blogger profile is empty :P
ReplyDeleteI tend to agree with you about SCTP although I did not have to make a kernel change to get started. It seems to work in vanilla Ubuntu now.
If you have any thoughts on more legit competitors to SPDY, I'd be eager to hear your thoughts. I've already investigated FireFox's CWND sorting. Anything else?
You should check out SDPY-over-SCTP by a prof at UDel:
ReplyDeletehttp://www.cis.udel.edu/~leighton/SPDY.html
He also did a HTTP-over-SCTP test with side-by-side video of pages loading under different loss ratios. (Though this HTTP also assumed a satellite-internet-level RTT, unlike the SCTP test.)