|
Grex > Coop9 > #9: A Scheme to Partition Grex |  |
|
| Author |
Message |
| 25 new of 121 responses total. |
scg
|
|
response 25 of 121:
|
Nov 24 19:02 UTC 1996 |
grumble...
|
dpc
|
|
response 26 of 121:
|
Nov 24 19:27 UTC 1996 |
grease?
|
robh
|
|
response 27 of 121:
|
Nov 24 20:20 UTC 1996 |
"gridlock"
Tell me that name won't be appropriate within its first ten minutes.
|
janc
|
|
response 28 of 121:
|
Nov 24 21:03 UTC 1996 |
(My suggestion of the name "edna" was 75% a joke. I have no terribly
bright ideas on this and certainly have no attachment to "edna". Calling
it "mail.cyberspace.org" would make a compelling amount of sense, perhaps
too much to be really Grexian.)
(I don't know how the going-to-another-machine-to-read-mail thing could be
made transparent. Maybe there is some way. Certainly it is possible, but
one of the goals of this plan is to come up with something that can be
implemented with existing parts and doesn't require really dramatic new
software. We might well be able to evolve toward that.)
Membership perks: If we allowed all users to telnet from Grex to Edna,
then many would telnet into Grex then from there to Edna to read their mail.
This would me (1) mail users consuming Grex's incoming telnet ports, and
(2) mail being sent over Grex's link as people read their mail. This this
would violate the idea of resource separation between Grex's mail and other
activities.
As currently configured, this would not be a major problem because ordinary
users can't telnet anywhere. However, as currently configured, members would
be able to telnet from Grex to Edna. This is a perk, but a fairly small one.
We could certainly configure the system so even members can't do this. But
Let's look at the alternatives:
(a) Allow Grex->Edna telnet to all users. This pretty seriously undermines
the whole idea of the partition. It's a big problem.
(b) Allow Grex->Edna telnet to members only. Minimal impact on the
partition, a noticable, but not huge benefit to members.
(c) Allow no Grex->Edna telnet to anyone.
The problem is that alternative (a) is the one that is a big problem. It goes
a good way toward overturning the whole scheme. Either (b) or (c) is
technically OK. So why should we prefer (c)? We'd have to write new patches
to the kernal to specifically to deny members the access to Edna (while
allowing them to telnet to any other machine on the internet). It seems
strange to bend over backwards to deny priveleges to members just for the
philosophy of it. Yes, I think we should prefer giving priveleges to all
users over giving them to just members, but I think it is OK to give
perks to just members when they don't cost us much of anything and the
alternative is to give them to nobody at all.
Still, if people prefer (c), that's fine. It's certainly do-able and
compatible with the rest of the scheme.
The question with POP-clients is similar. Right now, I think members and
only members can run POP-clients on Grex. Nobody does, and there aren't any
installed, but anyone could install one (I think). If we had a separate
mail machine, interest in such things would step up. We have the same
three alternatives, with the same costs. Telnetting someplace and running
a mail program isn't really much different in terms of resource usage than
using a POP-client to connect to a POP-server on that machine. Within the
constraints of the partition plan, the viable alternatives would be either
to allow it only to members or allow it to nobody.
|
scg
|
|
response 29 of 121:
|
Nov 24 21:27 UTC 1996 |
If we really had two physically separate Internet connections, then allowing
people to telnet from Grex to gridlock (hey, I like that name...) would be
a problem. If we have grex and gridlock sharing a link, but with gridlock
having a PPP connection to something on Grex's ethernet with the speed turned
down to keep it from using more than its share of bandwidth, that problem goes
away. Telnetting from the outside world, to grex, to gridlock, wouldn't be
any faster (probably a bit slower, actually) than telnetting to gridlock
directly.
|
popcorn
|
|
response 30 of 121:
|
Nov 24 22:37 UTC 1996 |
(I like the name gridlock too.)
|
dang
|
|
response 31 of 121:
|
Nov 25 00:32 UTC 1996 |
(yeah, metoo)
|
janc
|
|
response 32 of 121:
|
Nov 25 02:24 UTC 1996 |
I think I see. My original proposal looked like this:
+----------+ conferencing link ___________
| GREX | --------------------------------------- ( )
+----------+ __) (__
| ( The )
| local network link ) Internet (
| (__ __)
+----------+ ) (
| EDNA | --------------------------------------- (___________)
+----------+ mail link
(Actually there are some routers in here, that I didn't draw in.) Here the
mail link might get jammed with traffic, but the link to the conferencing
system remains uneffected. But if we open up general access to the local
link connecting the two systems we create an additional fast path to the
mail machine, so mail load (at least as far as reading is concerned) starts
to creap onto the other link.
What Steve Gibbard suggested looks kind of like this:
+----------+ +--------+ internet link ___________
| GREX | -------- | ROUTER | -------------------- ( )
+----------+ +--------+ __) (__
| ( The )
| mail link (slow) ) Internet (
| (__ __)
+----------+ ) (
| GRIDLOCK | (___________)
+----------+
Here packets destined for the mail machine have to go through the artificially
slow mail link. Because of this bottleneck, mail can't consume more of the
internet link than the capacity of the mail link, so the rest of the
internet link bandwidth is effectively reserved for the conferencing machine.
This has lots of advantages, I think. We only need one connection to the
net, so we look more normal to the outside world and it probably costs less.
We can more simply adjust the speed of the mail link, giving us more control.
We can let all users telnet to the mail machine, because their usage would
still be limited by the same slow mail link. Allowing POP clients might
still be an issue if we don't want the CPU load from them on GREX, but that
may be less of an issue than network bandwidth.
Steve's plan does seem like a better plan, if it is technically feasible.
|
nephi
|
|
response 33 of 121:
|
Nov 25 02:36 UTC 1996 |
I don't see why Steve's idea wouldn't be feasible . . .
I like the name "gridlock", but think that we should put
"mail" in as a cname, to keep the confusion down.
I think that if we end up with "gridlock" that we should
allow pop access to it since it won't hurt anything else
at that point.
|
e4808mc
|
|
response 34 of 121:
|
Nov 25 05:25 UTC 1996 |
Ok, I tried reading slowly and carefully, and I still can't figure out the
>answer. Here's the question:
>What impact, if any, does partitioning, reducing ptys (whatever they are),
>and all the other techie stuff have on us *very* simple users who are on
>antique equipment (Mac plus, 1200 baud modem), can only use dialin because
>we have no other access to computers, love conferencing, and use a moderate
>amount of email, again because Grex is the only internet access that we have?
>
This is the same question I had in 144, but I'd like to know the impact of
these two schemes in a frame of reference that I can understand.
|
nephi
|
|
response 35 of 121:
|
Nov 25 06:02 UTC 1996 |
>What impact, if any, does partitioning,
For a dial-up user, partitioning will increase the speed of Grex
for you. It will mean that mail will no longer take up all of
Grex's resources, but only a predetermined percentage. It will
also mean (in it's current form) that you may have to telnet to
another computer in order for you to read your email.
> reducing ptys (whatever they are),
Well, reducing ptys also speeds Grex up. It limits the number of
people who can telnet into Grex at any given time, such that fewer
Grex resources can be used overall.
>and all the other techie stuff have on us *very* simple users who are
>on antique equipment (Mac plus, 1200 baud modem), can only use dialin
>because we have no other access to computers, love conferencing, and
>use a moderate amount of email, again because Grex is the only internet
>access that we have?
If you use Grex as your ISP (<shock>), then this should speed your
connection to the Internet from unbelievably pitiful to slightly
less unbelievably pitiful. 8^) It will only serve to speed your
conferencing.
|
arthurp
|
|
response 36 of 121:
|
Nov 25 06:37 UTC 1996 |
I was thinking grind.cyberspace.org. That's what it would do. I'm so far
behind all the time. :(
|
rcurl
|
|
response 37 of 121:
|
Nov 25 07:04 UTC 1996 |
My mail on CAEN is all handled by srvr5 though I log into a different machine.
I have no idea how all that works, but maybe its the same idea as being
considered here. (There's another potential name for the mail server: grvr!)
|
ajax
|
|
response 38 of 121:
|
Nov 25 09:41 UTC 1996 |
Re 37, one way they may do it is to use NFS or something to make
remote drives appear like local drives, another is to run a POP or
IMAP server that certain mail clients can access. Those options
have been discussed on Grex before. I'm about to enter a separate
item to discuss using an IMAP server.
Re 32, in the second diagram, if people typically log into Grex and
telnet to the mail host to read mail, the network bandwidth usage is
about the same as if they'd telnetted to the mail host directly.
However, they'll be consuming a pty on grex, which uses some of its
limited CPU/memory resources. Running telnet uses less than running
pine, but it would still use CPU resources, so letting everyone telnet
from the grex to the mail host would not be as clean a "partition" as
making people telnet directly to the mail machine.
The shared line idea is what I described in #10; one drawback mentioned
there is that it's easy to set a maximum speed for the net->mail or
net->grex link from the router, but not as easy to set a minimum speed,
unless you limit the speed of both connections from the router, which
leads to inefficient network utilization. So in the example in the
second diagram, if grex is using a huge amount of bandwidth, gridlock's
share would be less than what the "mail link (slow)" connection allows.
Also, if we get an ISDN router like the Ascend P50, which uses an
Ethernet interface, we'd probably need a separate router (like gryps)
to limit the bandwidth usage of the mail host. The second router
isn't reflected in the second diagram in #32.
|
davel
|
|
response 39 of 121:
|
Nov 25 11:25 UTC 1996 |
Um. Also, with the second diagram: if we slow down the link to gridlock, it
seems all too possible that it will just get pounded with more (but slower)
sendmail sessions - or with sendmail attempts-to-connect, if we limit the
number of sessions. In either case, the traffic on the main link increases.
I could well be wrong, but that's what comes to mind as I stare at that
diagram.
Also um. Some years ago, when system names were being discussed, the idea
was that some other word for a group (flock, herd, crowd, whatever) be used.
A number were offered at that point.
|
nestene
|
|
response 40 of 121:
|
Nov 25 12:56 UTC 1996 |
1) I dial in to grex exclusively; if telnet from grex to edna is not
allowed, how will I read my mail? I am a member, but local non-
members may share my concern.
2) Is it possible to partition ptys so that some are not open to
telnetters but may be used for screen or other local purposes?
|
janc
|
|
response 41 of 121:
|
Nov 25 16:08 UTC 1996 |
The suggestion for dial-in users was that when you first connect to Grex
you are immediately confronted with a menu that asks you if you want to
connect to Grex or to the mail machine. You select one, and then the
usual login: prompt appears. You could log off Grex at any time to
drop back to the menu and access the other machine. Note that you could
send mail from either machine, but you'd have to switch from Grex to Edna
to read mail.
Yes, Marcus is planning to further modify telnetd so it will reserve some
ptys for things like screen and more significantly, for a terminal server.
|
kerouac
|
|
response 42 of 121:
|
Nov 25 18:10 UTC 1996 |
"Gridlock" is too impersonal. The new computer needs an identity. It
needs a name. Why not call it "Valerie" in honor of Popcorn and all the
other Valeries who have used grex:
"Valerie.Cyberspace.org" Has a nice sound to it! :)
|
janc
|
|
response 43 of 121:
|
Nov 25 18:33 UTC 1996 |
I don't like Gridlock too much. Gridly would be more affectionate.
|
tsty
|
|
response 44 of 121:
|
Nov 25 18:38 UTC 1996 |
.. fron grep to egrep ... from grex to egrex ... the egress for email.
and when it is realllllly fassssssssst .... fgrex
|
kerouac
|
|
response 45 of 121:
|
Nov 25 18:56 UTC 1996 |
or it could be simply "grex2".."grex2.cyberspace.org" Kinda
boring but it would be a way of simplifying things.
|
albaugh
|
|
response 46 of 121:
|
Nov 25 20:05 UTC 1996 |
As a grex member, I would be seriously annoyed if I couldn't telnet from grex
to the mail machine. I want to be able to do a variety of things with
"one-stop shopping", such as being able to !mail from within picospan. If
I had to !telnet from within picospan, then mail, that would be OK, But if
I had to exit picospan, logoff grex, login to mail-machine, etc., *that*
would be lousy...
|
albaugh
|
|
response 47 of 121:
|
Nov 25 20:28 UTC 1996 |
One other thing: With separate machines, would users get to have a common
personal disk storage area across machines? Or would they be separate?
The latter would be painful...
|
dang
|
|
response 48 of 121:
|
Nov 25 21:17 UTC 1996 |
I would guess that home directories would be in common on both machines, sort
of like the UM. In that vein, UM uses IMAP clients, which would defeat the
purpose of the seperation, because the CPU/memory hog of the mailer (Pine is
the primary IMAP mailer) would still be on grex, and the link flood of mail
going over the link would still go through the grex link. Question: Would
it be possible to have the router send information to the origional off site
source of the login, rather than through grex if someone is telnetted from
grex? This would get rid of some of the link drain on grex. You would still
have the CPU and link drain, tho.
This item is linked to coop 9
|
janc
|
|
response 49 of 121:
|
Nov 26 05:24 UTC 1996 |
Probably directories would be separate. To make them common requires NFS,
and NFS apparantly isn't very secure.
(Note that security questions for Grex and organizations like universities
often work out differently. Most places would consider anything that let
strangers log into their machines a terrible security hole. We consider
"newuser" a feature. But on the other hand, most places know who their users
are and are less worried than Grex is about what people can do once are on
the system. I don't know the details of the problems with NFS, but there are
reasons why things that other places consider reasonably OK can be more
serious concerns on Grex.)
|