[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: the "cluster" system call (and file system type)
El Domingo 15 Julio 2001 21:29, escribiste:
> On 20010715 Jordi Polo wrote:
> >> I think what people hate is polluting /proc in separate subtrees, like
> >> putting things in /proc/sys/net/cluster, /proc/cpu/cluster, etc.
> >> I thing thigs can be done in an transparent way defining a fake
> >> clusterfs, with things like cluter_add_entry etc. that (by now) just
> >> call they equivalents in proc_xxx, rooted at /proc/cluster. If sometine
> >> you have to move the tree, well, just clone the proc_xxx functions to a
> >> new independent tree.
> >
> >As far as i can see you are going to do a /proc-like fs outside /proc so
> > as david want the same but in /proc i think that we can begin to think in
> > what will be inside that directory and later we can put it in /proc or
> > /cluster or whatever.
> >We are just discussing the minor thing we'll have to decide, when we have
> > all we want to be inside that directory we can just ask linus where he
> > wants it. Maybe is just a matter of taste :P
>
> That's what I wanted to say, I did not make it very clear. If all the info
> and control is to be in a file system, I see (I'm not a kernel hacker) two
> ways, devfs-like and proc-like. I only see using proc as a fast hack to
> start using and designing something that perhaps will need some special
> design or feature.
> But using /proc to start work, people can focus on what features want
> there.
Perfect, i don't mind if we spend the next 3 months just discussing the
architecture . Otherwise i think we'll realize about things too late (with a
lot of effort coding done) , so let it be /proc if you like and let's start
to talk about arch.
I like a lot your idea about mounting a Fs to make our node member of a
cluster. There we can have :
1.-Info about other nodes (maybe processes but more overload) there, things
that we'll need to make better decitions(node number -ip , cost to arrive
there, ¿cpu mem loads?
2.- Configuration of the cluster parameters in the /proc way, we mainly have
to make it as automatic as possible but let the administrator control for
fine tuning.
Sorry to be so generic, i have no time now, i'll explain longer tomorrow
Something that worries me is how will we find the processes (as they go away
from his node) we have 4 alternatives:
1.- a central server knows all : uggh , i don't like any weak point like that.
2.- We do a broadcast (maybe multicast .....) requesting that process.
3.- The info where the process lives (i don't thing a process will migrate a
lot )is updated in the node where the process was born . so we can just ask
the node that according to the cpid is the local node and there will be the
info . If that node is down we can use 1 or 2 . I like this alternative most.
This is like a cache, if the info is not in that node we do a broadcast that
is a more expensive operation.
4.- if the processes don't migrate a lot makes sense update the info in
everynode ?? i have no idea about this, surely in a single ethernet with hub
it makes sense (if we have choosen number 3), but i don't know about other
configurations.
In broadcast in lans that use routers and are connected with other lans
members of a cluster. the router can get that info from every lan so if
other lan make a broadcast request it doesn't have to broadcast it to the
other lan because he has the info. Or just leave a cache of the last request
so it doesn't have to make a request next time.
And we surely want a cache in everynode of the processes requested (as it may
be requested again) similar to the arp cache but this cache is not trown
awayas arp but when we go to a node requesting a process that is not there
(he migrates but our cache is outdated) that node say as it migrated so we
choose 1234 above and update our cache, if that node is down we use 1234 also
Feedback really wellcome . (written very late at night , i'm not sure if it's
all rubbish :P)
--
Jordi
Student of Spain
Linux-cluster: generic cluster infrastructure for Linux
Archive: http://mail.nl.linux.org/linux-cluster/