I also prefer a base + mod(rank)/n approach. My problem with the
random
port selection with retry is that you introduce delays in the startup
process as you build the tree. Depending on tree depth, this could
cause
debugger startup to be slow. I don't think 65536 is the correct 'n'
since
you then end up scattering PTP ports across the entire user port
range,
and also, as Daniel points out, because of the multiple debugger
instances
case. I was thinking 'n' might be 256 or maybe 512 since it's not very
likely a user would ever run that many tasks on a node. There's
still the
slight possibility of port collisions because of other applications,
so
I'd 'reserve' a few more ports above base+n to use for collisions.
I'm not entirely sure why the parent SDM which is trying to build
the tree
downwards needs to use a random port connection approach. I'm also
afraid
that if you make 'n' 65536, that you will accidentally connect to a
random
port from some other application and either get yourself hung because
whatever you connected to doesn't understand your protocol, or
worse, you
hang or crash the other aplication.
You can at least partially solve this by telling the SDMs what
'base' and
'n' (if not hardcoded) are when you start the SDMs. Each task picks
its
port number using the base + mod(rank)/n calculation. The proxy also
knows
what 'base' and 'n' are, as well as task rank, so as it builds the
routing_file it can fill in the port number for each SDM. Then as SDMs
build the tree, they try to connect to that port.
The question is what do you do if you have a collision on port? The
parent
SDM can try connecting to ports in the spare port range until it
connects
to the proper SDM.. You still have the problem of connecting to random
applications. Also, what's the timeout when you try to connect to a
port
where nothing is listening? Could it be long eneough to make startup
time
a problem?
Your handshake to validate that a legitimate SDM is connecting is
important, especially since whatever connects (wrong user's SDM,
malicious
user) to the SDM could get control of the debugger on some tasks in
the
application. If you were to build the tree bottom-up instead of top-
down,
then as long as an invalid connection could do no worse than send
bad data
upstream to the GUI and not grab control of the application, the
risk is
less.
I'm not sure I'd rely on users to pick 'base'. If two users pick the
same
base, or one that causes overlap of port numbers, then you still
have a
problem.
A simplistic way to pick base would be a calculation based on the
user's
uid number (uid # * something mod(256)?). The only other way I know
how to
solve the problem is what we did in DPCL, where we had what we
called a
'super daemon' that ran as root and handled the issue of starting
multiple
instances of DPCL daemons, but that gets a little complicated.
How do you tell when the tree is built? Is this by each parent SDM
keeping
track of how many child SDMs it started and counting responses then
reporting 'done' up the tree?
Dave
Daniel Felix Ferber <dfferber@xxxxxxxxxxxxxxxxxx>
Sent by: ptp-dev-bounces@xxxxxxxxxxx
08/28/2008 01:18 PM
Please respond to
Parallel Tools Platform general developers <ptp-dev@xxxxxxxxxxx>
To
Parallel Tools Platform general developers <ptp-dev@xxxxxxxxxxx>
cc
Subject
Re: [ptp-dev] Questions about PTP SDM debugger
Greg and Dave,
I think that Greg suggestion to launch SDM is reasonable. But are we
considering race conditions? A am afraid that this approach might
present several failure patterns depending on how long each sdm delays
to start.
For example: The servers and the master are started nearly at the same
time. All servers bind to a port as you described. The master receives
the routing file and starts connecting to children that on their turn
connect to grandchildren and so on. What happens if a children
delays to
start up for some reason? Its parent will try to connect (but the
children will not be listening yet) and the parent will try the next
ports, but will never try again the port that the children is actually
listening to. I saw this happening, and that is the reason why the
launcher is currently starting the master after the servers instead
the
opposite as described in the specification. I think other race
conditions might be possible.
There is another issue in the strategy to launch the sdm master. After
starting the sdm master, the launcher starts listening on a socket
where
sdm master is expected to connect. The port number is passed as
parameter to sdm master. However, it may happen that sdm starts faster
than the launcher creates the socket. The sdm master will try to
connect, and on failure try the next ports. This approach does not
make
sense in this situation, since the port number passed to sdm master is
guaranteed to be the port where the launcher is listening. Therefore,
sdm master should not try the next ports, but try the same port again.
Another concern: Does the handshake consider the job ID? There could
be
a scenario were two users start a debug sessions on the same machine
at
the same time. Then, one might connect to the SDM server of the other,
by accident, if the are listening on the same port range.
I agree that using a base port number is better than using a random
number for each process. I think it is enough that the base port
number
is pseudo-random. I would avoid using a fixed port number because that
would potentially cause port number collisions on two simultaneous
debugging. I understand that sdm servers will know to handle this
collision, but the start of sdm servers will take more time. By
choosing
the base port randomly, we reduce the probability of causing
collisions.
My comments about who should write the hostfile: I see Dave
concerns. I
really did not consider that the amount of data to be transmitted
would
become that large.
Couldn't we establish a standard file format to be used for all
debuggers? Then the file could be written by the proxy, regardless
which
debugger is being used. I don't have a really good idea for this issue
yet.
Best regards,
Daniel Felix Ferber
Greg Watson wrote:
Good, I'm glad we're in agreement :-). Daniel, do you have any
comments on this?
Regarding the port numbers, this is not how I had intended the
debugger startup to work, so I want to change this at some point. My
approach is as follows, but any other suggestions would be welcome.
1. The SDM servers are given a "base" port number. At startup, they
attempt to bind to this port. If that fails, they try to bind to
base_port+1 after waiting a short random period (this is to avoid
servers started on the same node from chasing each other up the port
numbers). An alternative to this would be to bind to ((base_port
+rank)%65536)+1024. A third alternative would be to use a pseudo
random number generator seeded by the rank.
2. When the SDM master receives the routing file, it can then
determine the location of it's children, so it attempts to connect to
each in turn using the same port generation mechanism as in #1.
3. Once the connection is established to the server, a handshake is
used to swap credentials, etc., then the routing file is sent. The
routing file could be successively pruned as it propagates up the
tree
to reduce bandwidth.
4. Once the server receives the routing file, it does the same as #2.
5. This continues until all connections have been established, or
there was a timeout or some other error.
Greg
On Aug 28, 2008, at 8:46 AM, Dave Wootton wrote:
Greg
I think the proxy should be responsible for building the routing
file, in
order to keep the traffic on the connection between the GUI and the
proxy
down. With the current approach, you are sending node information
across
the connection twice, once to populate the PTP runtime model, then a
second time to create the routing file on the nodes where the SDMs
are
running. I'm not sure what the message length for the messages
from the
proxy to the GUI are, but for the remote_file you have
strlen(task_index)
+ strlen(hostname) + strlen(port_number) + 3 bytes per node. In my
case
that's close to 20 bytes per task, minimum. With large numbers of
tasks,
this could be a lot of data, and since all of these interactions
between
the GUI, the proxy, and the SDMs are a serial process, they slow
down
debugger startup.
The down side to this is the need for each proxy to implement
support
for
each of unique debugger startup sequences it is willing to support,
where
you could end up with some proxies not supporting a debugger. If you
implement all of the code in the GUI resource manager side though,
I'm not
sure you don't have the same problem, where the RM needs to be
aware of
the details of both the debugger startup sequence and the details
of a
particular runtime environment/proxy.
The other question I have after seeing the contents of the routing
file
you generate is the generation of random port numbers. If you end up
actually using these port numbers, do you run the risk of
accidentally
using a port number reserved for some other application, unless you
block
out a range of port numbers and only use that range? Even if port
numbers
are up for grabs with no expectation of reserved port numbers, what
happens if something else is using your port number?
Dave
_______________________________________________
ptp-dev mailing list
ptp-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/ptp-dev
_______________________________________________
ptp-dev mailing list
ptp-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/ptp-dev
