[Server-devel] Network Provisioning

[John Watlington]

On Apr 25, 2008, at 8:29 AM, Stefan Reitz wrote:

> Hi Y'all,
>
> On Fri, 2008-04-25 at 14:57 +1200, Martin Langhoff wrote:
>>
>> On Fri, Apr 25, 2008 at 2:38 PM, John Watlington  
>> <wad at media.mit.edu> wrote:
>> >
>> > Proposed change to the hardware spec:
>> >
>> > From one to four access points may use an simpler switch,
>> > connected to the server over a 100 Mb/s link. From five to seven
>> > access points will need a better switch, which provides a 1 GB/s
>> > link to the server.
>> >
>> >  This means that a 1 GB/s interface should be specified for the  
>> servers.
>
> I feel teleported back to my 3-19-08 mail:
>
> [...]
> (I am thinking  server clusters - red hat cluster suite is offering  
> some nice tools which I never got a chance to use / try - thinking  
> 3 servers with fail-over and increased performance for clients  
> (like two servers actually doing something...) would be a starting  
> point)
> Birmingham is looking at 49 schools with a total 14,000+ students.
> [...]
>
>>
>> Theoretically, yes... but perhaps this is a bit over the top. For the
>> space we are aiming...
>
> please define our aim

Martin was correct in that the aim of this discussion is rural  
schools in
Peru.   Birmingham, NYC, and others will require heftier servers.

I do dispute any claim that 1Gb/s network interfaces are over the top at
this point in time.   The cost difference on the manufacturing side is
around $2.

>> - the XS services will bottleneck well before saturating 1Gb/s  
>> traffic
>> - 'upstream' services that the XS is routing will bottleneck well  
>> before 1Gb/s
>>
>> if we see a 7-AP setup, it will be there to support either a large
>> number of laptops or a location with obstacles that needs many
>> antennaes. In any case, it will support laptops mostly peering w
>> eachother.
>
> how about those 14 - 28 AP setups?
>
>>
>> If we are designing for a "client base" of laptops that we actually
>> expect to saturate 1Gb, then... we need to start recommending a
>> mid-range server cluster, perhaps a SAN, all costing a few megabucks
> -->   http://en.wikipedia.org/wiki/Weasel_words
>>
>> ;-)
>>
>
> You get pretty decent off-the-shelf hardware for $500 (asus m2n,  
> athlon x2 xx , 2GB RAM) + cost of storage space) per server.
> And considering not every deployment is a remote underelectrified  
> mountain / desert / ... area, this should not be light heartedly  
> dismissed.
>
>> cheers,
>>
>> m
>
> And I agree with Aaron Huslage that the nature of the AP is going  
> to be another big hitter. But I really haven't seen the numbers on  
> the budget yet. A decent (non-WRT54(...)) AP comes for $300 - 450  
> and is still worth considering.


Please define decent, and how it differs from "cheap".   We have run  
into APs that appeared to artificially cap the number of connections  
to less than 30 (market segmentation ?), but have also tested $50 APs  
which seem to support 50 users fine.

Yes, centrally managed networks of APs are much better.
Is that the $300-$450 price you quoted, and does that include the  
controller ?

wad