LinuxTuples, a tuple space server for Linux
Will Ware
wware@alum.mit.edu
LinuxTuples is an open-source tuple space server, with associated code
for writing clients, designed to run on a networked cluster of Red Hat
Linux 8.0 boxes with x86 processors. The tuple space is maintained and
served from one machine on the the network. Tuple space is an elegant
and intuitive message-passing scheme to coordinate parallel
computations.
*******************************************************************
QUICK START GUIDE
You've already unpacked the tarball, so type "make". That will build
the executables tuple_server, tuple_client, and fft, and the shared
object library linuxtuples.so.
Start the tuple server by typing "./tuple_server 25000" on one
machine. On my cluster, that machine is "desktop", and 25000 is the
port number for the tuple space service. On all the machines in your
cluster, set these environment variables:
export LINUXTUPLES_HOST=desktop
export LINUXTUPLES_PORT=25000
Tuple space is a database of persistent tuples. Imagine a library,
where people can put in a book, get a book to take away, or read a
book without removing it so that others can still read it.
Because it's quick and painless to whip up tuple clients in Python,
and because Python is easy to read and understand, the first wave of
examples will be in Python. Later we'll look at C clients. The first
example Python code will print a dump of all the tuples in the tuple
space.
[wware@desktop linuxtuples]$ python
Python 2.2.1 (#1, Aug 30 2002, 12:15:30)
[GCC 3.2 20020822 (Red Hat Linux Rawhide 3.2-4)] on linux2
Type "help", "copyright", "credits" or "license" for more
information.
>>> import linuxtuples
>>> conn = linuxtuples.connect()
>>> conn.dump()
[]
That empty list means that the tuple server has no tuples yet. When we
write "linuxtuples.connect()", we are making a client connection to
the tuple server. The connection is a Python object with the following
methods:
conn.put(tuple)
puts a tuple into the tuple space
conn.get(template)
gets a tuple from the tuple space, which must match
the template, where None values are wildcards; blocks
until a matching tuple is found
conn.read(template)
reads a tuple from tuple space without removing it,
blocks until a matching tuple is found
conn.get_nonblocking(template)
non-blocking version of get(), returns None if no
matching tuple is found
conn.read_nonblocking(template)
non-blocking version of read(), returns None if no
matching tuple is found
conn.dump()
return the contents of the tuple space; if given a
list of templates, return only those tuples that
match at least one template
conn.count()
return the number of tuples in the tuple space; if
given a list of templates, return only the count of
those tuples that match at least one template
conn.log()
print a running log of tuple server activity to
stdout; use this in a "while 1:" loop
An alternate syntax for the linuxtuples.connect function allows you to
specify the host and port number explicitly, rather than taking them
from the environment variables LINUXTUPLES_HOST and LINUXTUPLES_PORT.
>>> import linuxtuples
>>> host = "desktop"
>>> port = 20000
>>> conn = linuxtuples.connect(host, port)
>>> conn.put((123, 3.1416, "abc"))
>>> conn.dump()
[(123, 3.1416, "abc")]
>>> conn.count()
1
Now we'll set up a client that performs some computational work. On
another machine, go into Python and type the following.
##########################
import linuxtuples
conn = linuxtuples.connect() # use args here, or env. variables
while 1:
# Handle requests to perform square roots
request = conn.get(("sqrt", None))
x = request[1]
conn.put(("sqrt done", x, x ** .5))
##########################
This will set up a worker task that performs square roots. Open up
another Python interpreter, on the same machine or another on the
network, and exercise the square root worker task with this squart
root client.
######################
import linuxtuples
conn = linuxtuples.connect()
conn.put(("sqrt", 5.0))
conn.put(("sqrt", 6.0))
conn.put(("sqrt", 7.0))
print conn.get(("sqrt done", 5.0, None))
print conn.get(("sqrt done", 6.0, None))
print conn.get(("sqrt done", 7.0, None))
######################
Here is what's going on.
(1) The worker task opens a connection to the tuple server and asks
for a tuple that matches ("sqrt", None), that is, a tuple whose first
element is "sqrt" and whose second element is unspecified. The worker
waits until such a tuple becomes available.
(2) The client task opens a connection and puts the tuple ("sqrt",
5.0). This matches the worker's request, so the worker removes it from
the tuple space and assigns it as the value of the "request" variable.
The worker then assigns x to 5.0 and puts the tuple ("sqrt done", 5.0,
2.23606) in the tuple space.
(3) The same client-worker dance happens with the 6.0 and 7.0 tuples.
(4) The client does a series of get operations, picking up the result
tuples created by the worker, removing them from the space and
printing them.
You can watch all this activity in amazingly gory detail with the
following Python code:
#############################
import linuxtuples
conn = linuxtuples.connect()
while 1:
conn.log()
#############################
*******************************************************************
Maintaining your Cluster
We don't want to do a lot of baby-sitting of individual computers in
the cluster. In the ideal case we should do something simple to get
them started, and then do all our work on one machine in such a way
that subtasks are automatically picked up by the other machines. It
would also be nice to do this in a way that gives reasonably even
load-balancing, where each machine gets roughly the same amount of
work.
Gelernter's solution to this was to invent a tuple operation called
"eval" which spawned worker tasks elsewhere in the system. That
approach strikes me as somewhat inelegant; and as described, it does
not address the problem of getting the software out to the other
machines.
I will assume that you'll use NFS to make executables available to all
machines, and that all machines will find them at the same directory.
I do something very simple on my cluster; I use NFS to map /home/wware
on one machine to all the /home/wware mount points on all the slave
machines, so that all the paths in my home directory are identical on
all machines.
(Note to self: It would be much better to use ssh and scp rather than
NFS. If you tarball the .ssh directory from one machine and clone it
on all the other machines, they all ssh to each other very happily.
The firewall setup for NFS is a bit complicated and it makes each
machine insecure.)
You can control jobs from a single computer using the jobcontrol.py
script. The various commands look like this.
jobcontrol.py log -- show a continuous running log of tuple
server activity
jobcontrol.py dump -- display the current contents of tuple
space; tuple elements truncated for
brevity when longer than 20 characters
jobcontrol.py size -- tell how many tuples are in tuple space
jobcontrol.py jobs -- tell what jobs are currently running on
the various machines in the cluster
jobcontrol.py empty -- empty all tuples from tuple space
jobcontrol.py start <program> <N> -- start N instances of a
program distributed around the cluster
jobcontrol.py stop -- stop all programs distributed around
the cluster
jobcontrol.py test -- run some tests to verify that the tuple
server is alive
You'll need to edit the 'slaves' list in jobcontrol.py, which is
currently set up for my cluster.
*******************************************************************
Thinking In Parallel
Many worker tasks will take the same general form as the square root
worker above:
while 1:
get a request
do a bunch of work
put the result
Square-rooting one floating-point number is too trivial a task to make
this worthwhile. The overhead for network communication overwhelmingly
outweighs any performance benefits from parallelism. Operations like
large matrix multiplies or inverses, or large FFTs, are worth doing.
In those cases, shipping tuples around takes a lot less time than the
actual work.
Suppose you have a bunch of workers all waiting for requests. The
efficient way to use them is to send out several requests at the same
time, and then collect the results. If you wait for the N-th result
before sending out the (N+1)-th request, you aren't taking any
advantage of parallelism.
When we put a request out into tuple space, there may be many other
requests of the same sort floating around, so we want to make sure we
get back the right result. We want to put a unique mark on the request
that we can use to identify the correct result.
There are two ways to do this. One is to use the input data in the
request as a field to be matched by the result. This is what we did
with ("sqrt", 5.0) and ("sqrt done", 5.0, 2.23606).
The other approach is to generate a few random numbers and use them to
key both the request and the result. Using three 32-bit numbers should
make it extremely unlikely that two different results will have the
same numbers, and be mistaken for one another. Four 32-bit numbers is
even safer.
It's generally preferable to use the input data as a key, if the input
data doesn't take up too many bytes. If it's much more than a few
dozen bytes, use some random integers instead. This will conserve
network bandwidth, as well as memory on the server.
There are a few operations that appear in the popular MPI
message-passing system that are obviously useful, and for these I
would like to present tuple-space equivalent operations. One idea in
MPI is that of a "barrier", a point in the code where different
programs wait for one another until all have reached that point. Then
they proceed with the rest of their computation.
MPI assumes a computational model where all the computers in the
cluster are running the same program, so it's meaningful to talk about
them arriving all at the same point in the source code. There is no
fundamental requirement in a tuple system that all programs would have
the same source code, but the notion of synchronization is still a
possibility.
One program will initiate the synchronization by putting a starting
tuple. This program needs to know how many programs are trying to
synchronize. Let's suppose it's three programs. The initial tuple
would then be ("synchronize", 3). The first element would need to be
unique enough to avoid collisions with other groups of programs that
might also want to synchronize, but that's another problem.
Each of the first three programs (which might include the initiator)
runs a piece of code like this.
str, count = conn.get(("synchronize", None))
count = count - 1
conn.put((str, count))
conn.read((str, 0))
As each program arrives in the "waiting area", it announces its
presence by decrementing the count of the synchronization tuple. All
the programs wait for the count to reach zero before moving on. The
initiating program should clean up the tuple space by getting the
("synchronize", 0) tuple. If it's left behind, it may cause confusion
with a future attempt to synchronize.
In MPI, there's a handy operation called "broadcast" which gets a
bunch of programs to agree on a value. This might be done, for
instance, to communicate the value of C's argv[] array from one master
task to several slave tasks. An example of the broadcast function's
usage is this:
MPI_Bcast(&n, 1, MPI_INT, 0, MPI_COMM_WORLD);
The master task assigns a value to the integer n, and that value is
written into all the slaves' n variables. The machinery for doing this
is hidden in the bowels of MPICH.
Broadcasting is easy to do with tuples. The owner of n's original
value puts a tuple like ("value of n", 17), and the other tasks read
("value of n", None) to learn the value.
Another sort of broadcast (for which, as far as I know in MPI, there
is no equivalent) is "public information", like the weather in Houston
or the current Dow Jones average. The owner of this information might
put a tuple periodically. When the value changes, he gets the old
tuple and puts a new one. Anybody wanting to know the value can simply
read ("Dow Jones average", None) and get a fresh value, without being
aware of when or how the updates occur.
The last parallel programming technique that I want to steal from MPI
is called scatter-gather. The idea is that you have a list of inputs
and you want to run some function on them to get a list of outputs,
and you want to let each machine work on a sublist.
MPI has a couple of tricky clever functions that will chop up the
input list into roughly-equal-length sublists, distribute them to the
machines, fetch the output sublists, and reassemble the entire output
list. This is all done with a couple of function calls.
A nearly-as-elegant approach is feasible with tuples.
n = len(input_list)
for i in range(n):
conn.put(("please process", i, input_list[i]))
output_list = [ ]
for i in range(n):
x = conn.get(("here you go", i, None))
output_list.append(x[2])
This assumes that we have hungry waiting worker bees all ready to turn
("please process", index, data) tuples into ("here you go", index,
result) tuples. But like MPI, we get nice even load-balancing with
this approach. Each worker bee tries to grab another request tuple as
soon as he finishes with the last one, so all the worker bees stay
busy as much as they can. Putting the index in both the request and
the response tuples makes it easy to reassemble the output list in the
proper order.
You might look at this and think, the master is lazy! He should do
some of the work himself! And in MPI, he does. But with this approach,
he will block on the get() calls, leaving the processor's time free
for the worker bee who will do some of the work. As long as each
computer in the cluster has at least one worker bee running, the work
will be distributed efficiently.
*******************************************************************
C Tuple Clients
You can exercise some C clients by starting the "fft" and
"tuple_client" executables. You'll see print-outs of how long it takes
to perform an 8K FFT. These programs don't print anything interesting
or instructive, but their source code shows how to work with tuple
space in C.
A "struct context" (defined in tuple.h) is equivalent to the "conn"
object we've been using in Python. It represents the information we
need to establish a connection with the tuple server. We don't
actually open a socket until we perform a tuple operation
(put_tuple(), get_tuple(), read_tuple(), get_nb_tuple(),
read_nb_tuple()) and then we'll close the socket before completing the
operation. The context struct remembers the server's name and port
number.
get_server_portnumber() takes a pointer to a context struct. It tries
to get the server name and port number from environment variables. If
it fails, we'll need to get them from command line arguments.
The make_tuple() function uses printf-style variable arguments,
starting with a format string, and returns a tuple. This isn't a
Python tuple, it's a LinuxTuples tuple, a data structure we will learn
about shortly. The possible elements of a tuple are indicated by
characters in the format string: "i" for a C int, "d" for a C double,
"s" or "s#" for a string, and "?" for a wildcard.
Strings can either be ASCII strings like "fft" or "fft done", or they
can function as buffers used for arrays or large data structures. This
is an efficient way to ship large blocks of data around.
Until I get around to writing more, study: fft.c, tuple_client.c.
*******************************************************************
C Data Structures
tuple.h
struct element
struct tuple
struct context
If you type "make htmldocs", and if you have doxygen installed on your
machine (it comes automatically with Red Hat 7.3 or later), then
doxygen will create documentation for these structures. Otherwise
you can just read the stuff in tuple.h.
*******************************************************************
External access to your tuple space
You may wish to access your tuple server over the Internet. Obviously
security will be a concern. To address this, you should access your
tuple server using SSH port-forwarding, where a port on your local
machine is re-mapped to a port on a remote machine, and the remote
machine trusts you because you have a copy of an SSH key from it.
Here are some links with info about SSH port-forwarding.
http://www.acl.lanl.gov/users/technotes/ssh_portforwarding.html
http://www.eskimo.com/support/ssh-forwarding.html
http://www.linuxjournal.com/article.php?sid=5462
http://www.yak.net/fqa/81.html
http://www.onlamp.com/lpt/a/1670
The onlamp.com page has something particularly interesting. I don't
ordinarily run the tuple server on the same machine that provides
SSH access to my house. So I need to route the SSH pipe thru my
house's SSH server to the machine where the tuple server is actually
running. The appropriate command is this:
ssh -f -N -L25000:desktop:25000 willware.net
The "-f" option means that SSH will ask for a password, and then fork
into the background, popping out only when the port-forwarding
connection is broken. The "-N" option means that we will not be
opening up a shell on willware.net, nor will we run any commands. The
"-L25000:desktop:25000" option means that when a program running on
the remote client sends a packet to its own port 25000, that packet
is actually delivered to port 25000 on desktop, the machine that runs
my tuple server.
To use the tuple server from another machine, for instance the server
at work, I type that command, and I also type:
export LINUXTUPLES_HOST=localhost
export LINUXTUPLES_PORT=25000
Then I can run a tuple space client on the work server. The connection
will close when the client is finished, and I'll need to type the long
"ssh" command again to reconnect.
This all assumes that the remote client (in my case, the server at
work) is trusted by the SSH server, and the way you do that is by giving
yourself an account on the SSH server and copying the remote client's
~/.ssh/id_dsa.pub file to the end of the SSH server's ~/.ssh/authorized_keys
file.
****************************** TODO *********************************
It might be nice to make up a GUI app that gave some kind of useful
snapshot of tuple space activity. Minimally this would be statistics
like the number of tuples in the space. There might also be some kind
of visual display of the tuples themselves, or maybe a colored pie
chart showing the proportions of tuples that match different criteria.
Here's what I'm doing to keep the C code tidy:
indent -i8 -br -npcs *.c *.h