Re: [DNSOP] TCP/UDP performances

Daniel Migault wrote:
> Hi,

Hi,

> We are looking for measurements on the following points :
>    - How TCP affects DNS servers performances compared to UDP?
>    - Proportion of clients that switch to TCP?

The proportion of clients switching to TCP depends on what are the
clients asking and their EDNS signaling. The protocol indicates if a
response comes with the TC bit on, the query should be retried using TCP.

To get a Truncated Response I've seen two scenarios:

1. The client doesn't support EDNS and the response won't fit on 512 bytes.
2. The client does support EDNS but announces a small buffer size
(usually 512 bytes), creating the same situation described above. A
general version would be a EDNS-capable client that announces a buffer
not big enough to hold the response.

>    - What kind of client are they? Are they those that do not implement
> EDNS0?

I mentioned two scenarios above, with or without EDNS. May be someone
else can think about additional conditions.

> 
> Feel free to provide any paper links you know.

I'd like to suggest you two documents to read. One is the Root Zone
Augmentation and Impact Analysis[1] from Duane Wessels and Geoff Sisson
and recent Duane's presentation at OARC[2] in Beijing.

[1]
http://www.icann.org/en/topics/ssr/root-zone-augementation-analysis-17sep09-en.pdf
[2] https://www.dns-oarc.net/files/workshop-200911/Duane_Wessels.pdf

Personally I've been working on a DNS benchmarking tool (like queryperf)
that retries queries when the response has the TC bit on. I can't share
any results because my testing hasn't been rigorous and the reduction in
performance is related to the fraction of queries you have to retry.
Also the benchmark can be affected if you re-use the TCP socket or if
it's created on each retry (and destroyed afterwards).

Cheers
Sebastian

> 
> Regards,
> 
> Daniel
> 
> -- 
> Daniel Migault
> Orange Labs -- Security
> +33 6 70 72 69 58
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> DNSOP mailing list
> DNSOP@ietf.org
> https://www.ietf.org/mailman/listinfo/dnsop