|
Just got back and put my computers together at home.
I've been leaving at 7am and not getting back till 9-10pm.
I thought the talk went well. It was standing room only with
about 150-200 people (I counted chairs afterwards).
I designed the talk so that people set up
the LVS and get it working. (Wensong, you have the talk
as I gave it, can you put it on the webpage? A few people
asked if they could get it) I thought
this was the most important thing since the project was
relatively new and I wanted people to be able to go
home and install it and get it to work, so spreading
the word around.
I brought a portable LVS with me (photos coming
as soon as they come off the camera and I find a scanner).
I was the first talk after lunch and had time to setup.
There were plenty of people who came early and just happened
to watch the setup.
I thought people would be more willing to believe
that the LVS worked if I was prepared to put my money
where my mouth was and could set the thing up while they were
waiting. I gave my talk
from the html on the 2 servers (one by tunnel and one
by localnode on the director).
I had planned to put markers into the html
on the two servers to show which server the page
was coming from, but I didn't have time for
that.
I explained the packet path twice for each setup
(NAT, TUN, DR), by showing the IP's and packet
path on a diagram,
the ippfvsadm commands and then showing a table
with the source and dest for the packet at each
hop and the contents at each stage.
At the end I pulled
one of the ethernet cables to demonstrate failover. The
mail from the alerts started pouring through every 20secs
and was clearly visible to the audience (murmurings etc
from the darkened room). The failed server
didn't get switched out (according to ippfvsadm),
I think because I had disabled
that part of the alert to test it when I'd set it
up a few days earlier. This was the only thing that went
wrong, but people seemed prepared to accept that the LVS
was responding to the cable being pulled.
There were no questions during the talk, which I interpreted
as everyone following the talk.
The talk took about 40mins.
There were a few questions afterwards -
What happens to your connection if a server fails (you loose
the whole connection and have to start your ftp again
or click on the link again)
Will the failed server be brought back into the LVS
when it comes back up again (yes, automatically
by the same script that took it out).
Some poeple wanted the tunneling explained again and I asked
them to come up after the talk (about 10-15 people did and
they seemed to get it eventually - the questions stopped
anyhow).
I brought buttons with Spike's logo on it to hand out
to people to wear. I was going to give them out
to anyone who'd installed LVS or was prepared to
answer questions about LVS at the Expo.
I brought 20, being all I could arrange in the time.
Knowing how fast people snap up anything free I was
dreading the disappointed people standing around waiting
for theirs. As it turns out only 4 people asked for them
after the talk and I have about 10 left. I'll reserve
some for the developers in the credits in the talk
otherwise anyone on this mailing list can have 1
for free on a first-come, first-served basis. Please
send me your s-mail address if I don't have it already.
There were some people in the audience who were well
aquainted with the theory. One is on this mailing list
and hasn't posted anything, so I'll let him remain
anonymous until he's ready to say something. He's either
using LVS for a project he has at work or is waiting
for the 2.2.x version to use in this project.
One person's comment afterwards was that the LVS
is a level 4 switch, which I hadn't thought about.
One person was interested in the geographic
based load director and wanted me to let him
know when it was done.
The main impression I got was that people hadn't installed
it and didn't know much about it. I assume people had
read the write-up in the conference procedings and had
recognised it as being useful enough to them to come
to the talk.
Since people in the audience didn't know much
about the project, I decided not to mention the
problems we were having with PHT unless anyone asked.
I had a little speech ready if any questions came
up but none did. PHT came up at the ExtremeLinux
BOF where PHT was mentioned as providing a Cluster
Server based on the LVS source patch for what they
thought would be $1000/server. Some of the
people in the BOF had already told me that they weren;t
happy about the credit PHT was giving LVS for it's work.
At $1000/server I didn't think the rest of the people in
the room would be interested in PHT's product.
I was asked for comments and decided to say nothing.
Perhaps I should have - can't tell if it would have
been better to or just different. With the audience
at the LVS talk, I had decided that the best approach
for the first public revealing of the code, would
be to get as many people to install it as possible.
PHT was in the preliminary schedule to give a
talk earlier the same morning, but wasn't in
the final schedules. I didn't see PHT anywhere
at the conference.
Setting up the server demo didn't go real well
and pointed out to me the pitfalls of relying
on other people's hardware. I originally
thought I was going to have a mobile 4 node
Beowulf to use. Although I allowed plenty of
time to set it up (a whole weekend)
Robert Brown from the
ExtremeLinux group, who kindly offered me
the setup, was too busy with his own work
for me to be able to get in to set it up. As it turns
out he'd just bought the 4 nodes for himself
and first one then another node wouldn't boot.
I was left the night before the talk without
a setup. I brought in my rather heavy 2 node
beowulf, which was already setup and working,
and was going to use one of his 2 working nodes as a client.
Still I had problems. I was using coax and
their machines were 100BaseT. I borrowed
2 100BaseT NICs (realtech 8139) and eventually got a NIC
to work in one machine, but the mouse module
wouldn't load at the same time, so no X there.
The same NIC driver loaded on my other
node (with indentical mobo, one node glibc-2.1
and the other libc5) but ifconfig gave
me "invalid flag errors". I got the same
thing with the driver in the kernel and with
another NIC (Intel 10/100) showing the problem
was not with the cards but my machine.
Someone suggested I
pull all the unneccessary cards from my
machine and it worked (I thought PCI cards
auto negotiated interrupts, but apparently
not in the 2.0.x kernels). The machine with 1 SCSI
card (ncr53...) wouldn't work with the NIC,
but the one with 2 SCSI cards was happier
(loaded the nic driver, but not the mouse
driver). I spent much
of the morning recompiling the kernel on a
100MHz pentium classic. This machine did load
the mouse driver and I could X with it.
When I got to the room I realised that the client
machine was setup for a 1600x1280 monitor which
the projector wouldn't handle. I couldn't change
the resolution with alt/ctrl/+-, but no problem, telnet
in to the client and DISPLAY back to the only machine
which could X. However I couldn't login as root
remotely to export the DISPLAY... fixed that.
After having started at
7:30am to set everything up it took till 1:28pm,
2 minutes before the talk, before I had the
client displaying onto the screen for the audience,
the html stored on the servers. I then found
that the failover was sending mail alerts
but not removing the failed server from the ippfvsadm
table. Steven Tweedie at this stage was trying to
find out who I was so he could introduce me.
It sounded like he was the person who put the talk
into the conference in the first place as he
knew a lot about the project and about Wensong.
At 1:30 I started the talk and decided
that I would have to go with an imperfect failover
demo.
I had brought a TV camera to record the talk
to send off to Wensong, but didn't have time
to set it up.
Matthew, would Rebel.com be up for a loaner
of 3 netwinders or some of those nice rack units
(2 or 4 machines) for a subsequent demo?
This was not the way to go.
BTW I saw Scott and gave him my business card.
Further publicity:
A local group, the North Carolina System Admins
(to which I belong) has asked for a demo
at their next meeting (2 weeks). Also my
work wants a talk too. (Matthew, any chance
of a few boxes for these, I'll ship them
back insured etc...).
The Linux J. wants an article. If it's OK
with everyone I suggest
Wensong write the main part and
I'll write a side bar about installing it.
They want no more than 3200 words (4 pages).
Wensong how about I talk to you offline.
Todo:
1. I'll write a HOWTO. It can be on the
LVS webpage for comments for a while and
when everyone is happy with it, I'll send
it off the the LDP.
2. It would be nice if the LVS webpage could
be on other machines and served
by LVS. I'll see if I can use my machine
at work. I will have to ask the right people,
but I think I can make a reasonable case.
It's a solaris machine and I'd have to use
a high number port. Anyone else got a site
they want to volunteer?
Thanks Matthew and Wensong for your help
with getting the talk going.
Joe
--
Joseph Mack, NA3T, FM05lw EME(B,D)
mailto:mack@xxxxxxxxxxx
AZ_PROJ map server at http://www.wm7d.net/azproj.shtml
|
