[c-nsp] Application Protocol Performance in low latency envrionments

[Tim Durack]

On Wed, Aug 26, 2009 at 6:23 AM, Mikael Abrahamsson <swmike at swm.pp.se>wrote:

> On Wed, 26 Aug 2009, Ash Net wrote:
>
>  The reason for performance degradation solely seems to be latency related
>> since there's tons of b/w available in the lab setup and over 10G lanphy
>> paths. Do people still deploy QOS for better traffic management on the
>> lanphy interfaces even with no saturation?.
>>
>
> All the protocols you mentioned are "query/response" ones and thus they
> take a big hit when latency is introduced.
>
> I know several companies who nowadays has a wan simulation device between
> the clients and servers in their dev labs, just so that the developers will
> develop applications that actually work in real life, not just in the 1/10
> ms latency of the dev environment. Imagine the difference in a latency
> environment between doing a single nested SQL query, as opposed to doing 1
> returning a list, and then doing one query per list entry. In a 1/10ms
> environment the difference might not be noticeable, but in a 100ms
> environment it most certainly will.
>
> At several ms network latency you're effectively dealing with harddrive
> latencies as opposed to almost memory latencies, and thus techniques that
> work akin to NCQ or alike needs to be employed. Failure to do so will create
> performance problems in real life.
>
>
Serious feelings of deja-vu. We lived through a similar scenario over the
last few months. GigE "WAN" (really a MAN) with 3-5ms latency. Had to
install some WAN accelerators, 'cos they promise to fix everything. WAN
Accelerators don't work in this scenario for the reason Mikael mentions (the
IO subsystem of a WAN accelerator is in the 2-3ms range.)

Only real answer is to get application developers to code for the Internet
at large, rather than for <1ms LAN. Of course our devs don't even like their
dev servers being 3-5ms away...

Tim:>