Skip to content
Find file
Fetching contributors…
Cannot retrieve contributors at this time
executable file 298 lines (239 sloc) 13.1 KB
(Copyright 2006 Sriram Srinivasan)
Kilim IFAQ: Infrequently Asked Questions. Kilim v 1.0
-- sriram srinivasan (Kilim _at_
Why is multi-threaded programming considered so hard?
It is relatively easy to get thread programming correct (to a first
approximation) by synchronizing all your shared data structures and
taking locks in the right order.
You could have one giant lock and just do things one at a time (like
the current python interpreter with its Global Interpreter Lock).
Clearly, this is not efficient. Increasing concurrent access of a
data structure (by using finer-grained locks) is what makes it
error-prone and hard to debug.
Kilim uses kernel threads. Where do tasks and threads meet?
Kilim's tasks are cooperatively scheduled on a kernel thread pool.
Tasks are needed when you want to split up your workflow into small
stages and write code as if it is blocking (instead of writing a
callback and having to jump to that function when it gets called).
Tasks should not ideally make thread-blocking calls, although if you
*have* to call one, it is not the end of the world. That's what other
threads are for .. they'll take care of the other tasks meanwhile.
A Kilim task is owned and managed by a scheduler, which manages the
thread pool. When a task needs to pause, it removes itself from the
thread by popping its call stack, remembering enough about each
activation frame in order to help rebuild the stack and resume, at a
later point). The scheduler then reuses that thread for some other
You can have more than one scheduler (read: thread pool) and assign
each task to a particular scheduler. See the bench directory for
How lightweight is "lightweight"?
The amount of memory occupied by a task is:
1. The java object that represents the task class
2. If paused, an array of activation frames is stored. The Kilim
weaver performs data flow and live variable and constant analysis
(intra-procedurally) to ensure that it capture only as
much as is needed to resume.
3. The contents of all mailboxes that the task is receiving on.
Clearly, all these depend on your application.
The depth of the task stack is limited only by the thread's stack; no
memory is preallocated. Note that when written in the message passing
style, stacks tend not to be too deep because each task is like a
stage in a workflow, with its own stack.
What's the difference between channels in Ada, CSP/Occam, Newsqueak,
Alef etc. and Kilim's mailboxes?
Most of these languages use synchronous channels as their basic
construct, where a sending task can proceed only after the receiver
has received (or vice-versa).
1. Synchronous channels are easier to reason about because there is
automatic flow control; the sender does not proceed unless the
recipient drains the channel. Tony Hoare, Robin Milner, Rob Pike
and John Reppy have all written extensively about synchronous
programming, so I will take their word for it. However, I still
find asynchronous programming (through buffering) a better default
choice for practical reasons:
2. Context switching has a cost, however inexpensive Kilim's tasks are
to create and context-switch (unlike the Occam/transputer world
with its hardware-assisted switching). Although Kilim's mailboxes
can be configured to be synchronous, it is not the default. There
are many cases where you want to send messages to multiple
recipients before waiting to collect replies. I find tedious the
CSP approach of spawning a task to avoid blocking while sending.
3. I like the same interface for both concurrent and distributed
programming (although support for distributed programming is yet to
be bundled with Kilim). Synchronous _distributed_ programming is
horribly inefficient .. every put has to be acked when a
corresponding _get_ is done.
This is why I have followed Erlang's example to prefer buffered
channels (called mailboxes) as the default choice.
Erlang vs. Kilim
Kilim is an ode to Erlang (, and strives to bring
some of its features into the more familiar Java world.
The term "Erlang", like Perl, refers to both the language and the sole
available implementation. Comparisons have to be made on these two
axes separately.
The Erlang language is a soft-typed, purely functional language and
has many of the goodies of a functional setting: higher-order
functions, beautifully simple syntax and pattern matching on terms,
features that I'd love to see in Java. However, programming in a purely
functional style is not everyone's cup of tea and there is no reason
that higher order functions and pattern matching can't be made
available in an imperative setting (See Scala, JMatch, Tom(from INRIA)
etc). If you have to have types, it is better to have Ocaml-style
types (or even Smalltalk); but compared to Java-style types, I prefer
the simplicity of Erlang's soft types.
The argument for Java lies not in the language, but in the incredible
JIT compilers, JDK, enormous open code base and community, excellent
IDEs, good network, database, GUI and systems support. Why throw away
all that?
The Erlang *environment* (not the language) offers lightweight
processes, fast messaging, uniform abstraction for concurrency and
distribution and many, many systemic features (process monitoring,
automatic restart), process isolation, failure isolation etc. These can be
built atop Kilim as well.
The idea behind Kilim is that one can have all the features of the
Erlang environment without having to move to the Erlang
Kilim vs. Transactional Memory
Hardware/Software Transactional Memory is currently the new hope and
an alternative for concurrent programing in the shared memory
world. It is appropriate in a mostly functional setting where most
objects are immutable and side-effects are rare or contained. In an
imperative setting, I have my doubts about TM's scalability; hotspots
are expensive. Atomic sections can't be too big, otherwise they risk
getting retried all over again. And the part of code that retries had
better not have any side effects that doesn't know about or is not
controlled by the TM, such as sending messages on the network.
I think the task and mailbox approach is a more understandable model,
has nice run-to-completion semantics, has convenient graphical
representations (dataflow diagrams, workflow diagrams, Petri nets). It
brings the interaction with other processes out in the open. It allows
batched and efficient communication.
That said, there is absolutely no reason not to use the TM facilities
internally inside Kilim. I intend to use non-blocking data structures
when they perform well (currently, Java's data structures aren't
as fast as I'd like them to be)
What's the relation between CCS/pi-calculus and Kilim
The notion that the Mailbox itself is a first class message datatype
and can be sent as part of a message is inspired by Prof. Robin
Milner's pi-calculus. This allows the topology to change with time.
A can send a mailbox in a message to B, B can forward that message to C
and C and D can shared that mailbox.
Beyond that, CCS, like CSP is a modeling and specification language,
and uses synchronous interaction between processes. At a practical
level, this is terribly inefficient (esp. in Java).
RMI vs. Kilim
We need to distinguish between RMI implementations and the concept.
RMI implementations block the java thread. That's a no-no for
scalability. They themselves are incredibly heavyweight -- I/O
serialization is always used, even in a concurrent setting, for
ensuring isolation. The request response paradigm doesn't allow many
other patterns of communication: fork/join, flow control, rate
control, timeouts, streaming etc.
Kilim, in a concurrent (local) setting, is at least 100x faster than
Java RMI on even the simplest benchmarks. In a distributed setting,
the Kilim approach is better because asynchronous messaging is much
more scalable. Combine this with automatic stack management and you
get a far easier programming model
What are Continuations and what is Continuation Passing Style(CPS)?
There is so much doubt and misinformation on the topic that a few
words are in order.
Simply put, a CPS style of programming is where a "return" keyword is
not needed.
The notions of procedures calling procedures by building up a stack
has been burnt into our collective programming consciousness. If a()
calls b() calls c(), we think, the stack must be three deep.
Suppose a() has nothing more to do after calling b(). It (that is, a()) really
doesn't need b() to return to it, so there is no use pushing a return
address on the stack. In other words, the flow of control _continues_
from a to b, never to return. Most respectable code generators
recognize this special case and prevent the stack from building up
("tail call optimization"). It is a pity this isn't available under
the standard JVMs. Even GCC doesn't do it all the time.
Now consider,
a() {
do stuff
do more stuff
b() {
Now you need a stack and you want b() to return in order to "do more
stuff". However, this bit of code can be transformed to ensure that b
doesn't return; instead it continues on to another procedure that
performs the "do more stuff" bit.
a() {
do stuff
b("c") // pass a reference to c()
b(nextProc) {
call nextProc
c() {
do more stuff
The "do more stuff" part has now been separated out into c(). Now,
a() chains on to b, supplying it the name of the next call to
make. For its part, b _continues_ to the procedure referred to by its
nextProc parameter, instead of returning.
This transformation ensures that you never need the "return" keyword
... you always continue onwards to the parameter supplied.
What if "do more stuff" needed to refer to local variables in a()'s
stack frame? Well, the transformation ensures that a() packages the
values of those variables along with a reference to the next proc to
call. Now, instead of "nextProc", we have an _object_ (with state and
a procedure) called a continuation.
The obvious question is, why bother? The stack worked well, didn't it?
Why dispense with it? Yes, the stack works incredibly well, which is
why CPUs and compilers have special support for it. However, the
continuation passing style allows for other forms of transfer of
control very simply. C++ and Java provide two forms of "return", one
normal and another using exceptions. If we had CPS, we wouldn't need
these special cases.
Instead of a() installing an exception handler, it would pass in two
continuation objects to b() that know what to do under normal and
under exceptional conditions. b() simply chains on to the appropriate
object as its last move.
As another example, you can have tasks that pass control to a
scheduler that in turn passes control to another task, all without
having to return to whoever called it.
In a programming language with explicit support for continuations (ML,
Lisp, Haskell), one can have the "return" keyword merely as a
syntactic sugar (like a macro). Internally, the compiler CPS
transforms the entire code, so no procedure returns to its caller.
Are there any disadvantages of continuations? Oh yes. Machines are
so well optimized for stack usage and no tail calls that the
system is biased against continuations, performance-wise. The
continuation object has to be allocated from the heap and depends
on garbage collection. This is one reason why OCaml doesn't use CPS.
That said, the current crop of garbage collectors and the amortized
cost of garbage collection often matches that of stack-based
allocation, and continuations are simply too powerful a feature to
Where does Kilim fit into all this?
Kilim's transformation is similar to CPS, but it needs to live within
a JVM that does not even support tail calls. It also needs to live
with the Java verifier that doesn't allow random gotos to be inserted
in the code willy-nilly. More details in the paper "A Thread of One's
Own" (included in the docs directory)
Something went wrong with that request. Please try again.