[Server-devel] Two separate sets of install scripts?

[Greg Smith (gregmsmi)]

Hi Wad et al,

Thanks for the info.

Tough problem but sounds like you are narrowing it down, good luck!

I hear you that radio network is a bottleneck more than CPU, memory or
disk on the school server. I would be surprised if PHP SQL calls to
MySQL were a bottle neck unless PHP is as bad as Ivan makes it sounds
:-)

We should still do some XS performance characterization if someone has
time. It can be as simple as a shell script running a bunch of WGET
commands on clients with output of TOP or PS from the XS.

On the network bottle neck, fortunately the Nepal use case will have
very light load:
http://wiki.laptop.org/index.php?title=Nepal:School_Server_Specification
#School_Server_Use_Cases

Activities are loaded on Xos and kids run them from there (BTW congrats
to Nepal team for the beautiful design and art in the activities:
http://www.olenepal.org/activities_download.html !). 

Any Mesh activity will be XO to XO. There's also no Nepali content on
the Internet so not a lot of surfing there. I can see man kids hitting
the Library Server simultaneously. If Squid - XS design works that
should all get served from school server.

Still, I think we need to solve this problem: "it is easy to saturate
three channels of mesh while testing the web filtering/caching with just
three or six laptops"

Some preliminary ideas:
Maybe a more powerful access point? A hardware solution may be too
expensive so SW/ network design is better but best is to have both
options. We need a way to characterize this and do some tests:

One test idea:
Associate Xos to the wireless AP routed to Squid/XS. In other words XS
doesn't need to be in the mesh:
Eg. XO <-> Wireless AP <-> Squid <-> XS or Wireless/AP depending on dest
IP <-> Internet.

XO - Wireless AP is over the air, everything else is hard wired.

Then compare that to the original suggestion where Mesh channels
(tunnels?) all traffic through XS to internet:
XO <--> XS <--> Squid <--> Wireless/AP

Or have both on at the same time so XO connects to mesh and wireless AP.
What happens then? 

I'm not asking you to do these tests, just trying to design some good
test and test environments. 

Its not clear that Tony or others would get meaningful results with only
one XO. If someone can figure what kind of HW we need to emulate
saturating the mesh I can try to procure it. I'll also start looking for
a wireless design specialist to comment.

Performance characterization is a bottomless pit :-(

Sulochan and Bryan,

We should plan to process your squid logs:
http://www.squid-cache.org/Scripts/. I assume you will remove all
identifying info, but it would be helpful to know how much traffic goes
out of the school, how many cache hits, most active times of day,
response time, concurrent connections etc. The logs will fill up your
server and be deleted so we want to process them and save that info
instead.

Also, please tell the students and teachers that everything they do can
be recorded. Hopefully it wont be, but they need to know that it can be!
That should be made clear in every deployment! Please make sure that
message gets across. Its hard to keep up with the political situation in
Nepal but I think its had some dramatic changes recently. The last thing
we need is someone capturing this traffic and using it against the
teachers or kids!

Non-technical point there but thinking about the log files made me
nervous.

FYI for all the bug system is at: http://dev.laptop.org/query 
There's a component for school server but only a handful of open bugs
listed.

I did see one where volunteers might help: security audit:
http://dev.laptop.org/ticket/1506 maybe someone could offload Wad, get
some port scan/audit tools and run that?

Too many open words and open questions :-( 

Read at your leisure and have a nice weekend!

Greg S

-----Original Message-----
From: John Watlington [mailto:wad at laptop.org] 
Sent: Thursday, February 14, 2008 2:25 PM
To: Greg Smith (gregmsmi)
Cc: John Watlington; server-devel at lists.laptop.org
Subject: Re: [Server-devel] Two separate sets of install scripts?


The failure point won't be the server, it will be the network.  Right
now we are rushing to fix problem in the collaborative middleware which
explode when a number of laptops sit on the mesh.  Just playing back a
single trace times 200 doesn't come anywhere close to emulating it.

Likewise, it is easy to saturate three channels of mesh while testing
the web filtering/caching with just three or six laptops...

wad

On Feb 14, 2008, at 12:57 PM, Greg Smith (gregmsmi) wrote:
 
> Hi Tony et al,
>
> I think Sulochan or Bryan need to confirm on these points.
>
> FYI I will be on the road most of the next three weeks so address any 
> specific questions directly to them. I'll watch and try to extract any

> consensus or decisions for posting in the wiki.
>
> BTW I may have recruited 1 - 3 more volunteers from the XO support 
> gang to build Nepal style XSX machines. I'll let them confirm when 
> ready. My goal is to document a very well defined Nepal implementation

> by the first week of March. Test beds can mirror that so any problems,

> bugs or new features can be reproduced and verified before being 
> deployed in production in Nepal. That may be optimistic but we'll 
> see...
>
> The hardest thing to reproduce may be 200 x Xos all squawking on the 
> Mesh at the same time. I wonder if we could capture an ethereal
> (http://www.ethereal.com/) or TCP dump of all traffic on an active XO.
> Then emulate the XO (QEMU?) on a high end server, generate the same 
> traffic (http://wiki.ethereal.com/Tools) and send it out at 200x 
> original data rate. That wont emulate 200 radios or 200 clients but it

> may help quantify the performance of the XSX under load (e.g. HTTP, 
> BW, CP/SCP over mount (NFS?) etc.) Just an idea, target phase 2.
>
> There's still room for more test beds so just reply here to get in the

> loop. There's no need to sign an SLA :-) You just need motivation, and

> a box.
>
> Thanks,
>
> Greg S
>