Papers by sriram srinivasan
This paper demonstrates an architecture for suspending and resuming methods in Java using a restr... more This paper demonstrates an architecture for suspending and resuming methods in Java using a restricted form of continuation passing style (CPS) transformation. It describes Kilim 1 , a toolkit to portably weave threads of control called Fibers, through Java code. The chief contributions of this paper are the set of design choices made for both space and time efficiency in getting one-shot continuations to work on the JVM (in some cases, 60x faster than competing approaches) and to address some tough issues traditionally passed over by others, such as handling of local subroutines and constructors. We are able to support hundreds of thousands of threads of control with switching times of the order of 3 to 4 μs on a low-powered laptop with Sun's JVM.
Kilim: A Server Framework with Lightweight Actors, Isolation Types & Zero-copy Messaging Sriram S... more Kilim: A Server Framework with Lightweight Actors, Isolation Types & Zero-copy Messaging Sriram Srinivasan Internet services are implemented as hierarchical aggregates of communicating components: networks of data centers, networks of clusters in a data center, connected servers in a cluster, and multiple virtual machines on a server machine, each containing several operating systems processes. This dissertation argues for extending this structure to the intra-process level, with networks of communicating actors. An actor is a single-threaded state machine with a private heap and a thread of its own. It communicates with other actors using well-defined and explicit messaging protocols. Actors must be light enough to comfortably match the inherent concurrency in the problem space, and to exploit all available parallelism. Our aims are two-fold: (a) treat SMP systems as they really are: distributed systems with eventual consistency, and (b) recognize from the outset that a server is always part of a larger collection of communicating components, thus eliminating the mindset mismatch between concurrent programming and distributed programming. Although the actor paradigm is by no means new, our design points are informed by drawing parallels between the macro and micro levels. As with components in a distributed system, we expect that actors must be isolatable in a number of ways: memory isolation, fault isolation, upgrade isolation, and execution isolation. The application should be able to have a say in actor placement and scheduling, and actors must be easily monitorable. To Alka, my dearest. Life's a wonderful journey, and I have the best travel companion. To Amma and Appa, for the easy laughter, love and encouragement, and the early mornings and late nights they have lavished on me. This dissertation is dedicated to Amma. She would have exclaimed "hai"! To my parents-in-law, for being an additional set of parents. One should be so lucky. To my advisors. Prof. Jean Bacon, for making this excursion possible, for allowing me to gradually discover my interests and strengths without pressure, and for her gentle and selfless guidance to collaborate with other groups. To Prof. Alan Mycroft, for setting distant Bézier control points that so significantly influenced the arc of my learning, for his exhortations and for his unwavering belief that I would get there. To Dr. Ken Moody, for his critical reading of my papers and this document, for his cheerful and enthusiastic cross-pollination of ideas, and tours of beers and wines. To Reto Kramer, for his friendship, guidance, technical knowledge, and a brain the size of a planet. There are a few dissertations lurking in there. To Lauri Pesonen and David Pearce, for the critical reviews, for the long coffee sessions packed with techie talk, and for sharing my enthusiasm for programming languages. To Eiko Yoneki, Pedro Brandão, David Eyers, Dave Evans, Minor Gordon and Niko Matsakis for sharing their impressive knowledge, opinions and good humor, for challenging my biases and assumptions, and for the gentle prodding to get on with it. To Boris Feigin and Kathryn Gray for introducing me to the world of types and semantics, and holding me accountable to rigor.
This paper describes Kilim, a framework that employs a combination of techniques to help create r... more This paper describes Kilim, a framework that employs a combination of techniques to help create robust, massively concurrent systems in mainstream languages such as Java: (i) ultra-lightweight, cooperatively-scheduled threads (actors), (ii) a message-passing framework (no shared memory, no locks) and (iii) isolation-aware messaging. Isolation is achieved by controlling the shape and ownership of mutable messages-they must not have internal aliases and can only be owned by a single actor at a time. We demonstrate a static analysis built around isolation type qualifiers to enforce these constraints. Kilim comfortably scales to handle hundreds of thousands of actors and messages on modest hardware. It is fast as well-task-switching is 1000x faster than Java threads and 60x faster than other lightweight tasking frameworks, and message-passing is 3x faster than Erlang (currently the gold standard for concurrency-oriented programming).
In addition, the book patiently explains all sorts of language details you've always wanted ... more In addition, the book patiently explains all sorts of language details you've always wanted to know more about, such as the use of references, trapping errors through the eval operator, non-blocking I/O, when closures are helpful, and using ties to trigger actions when data is ...
Workshop on New Horizons in Compilers, 2006
Abstract. This paper demonstrates an architecture for suspending and resuming methods in Java usi... more Abstract. This paper demonstrates an architecture for suspending and resuming methods in Java using a restricted form of continuation passing style (CPS) transformation. It describes Kilim1, a toolkit to portably weave threads of control called Fibers, through Java code. The ...
Internet services are implemented as hierarchical aggregates of communicating components: network... more Internet services are implemented as hierarchical aggregates of communicating components: networks of data centers, networks of clusters in a data center, connected servers in a cluster, and multiple virtual machines on a server machine, each containing several operating systems processes. This dissertation argues for extending this structure to the intra-process level, with networks of communicating actors. An actor is a single-threaded state machine with a private heap and a thread of its own. It communicates with other actors using well-defined and explicit messaging protocols. Actors must be light enough to comfortably match the inherent concurrency in the problem space, and to exploit all available parallelism. Our aims are two-fold: (a) to treat SMP systems as they really are: distributed systems with eventual consistency, and (b) recognize from the outset that a server is always part of a larger collection of communicating components, thus eliminating the mindset mismatch bet...
System and method for prepreparing a transaction process involving a chain of servers in a circular flow
This paper describes Kilim, a framework that employs a combination of techniques to help create r... more This paper describes Kilim, a framework that employs a combination of techniques to help create robust, massively concurrent systems in mainstream languages such as Java: (i) ultra-lightweight, cooperatively-scheduled threads (actors), (ii) a message-passing framework (no shared memory, no locks) and (iii) isolation-aware messaging. Isolation is achieved by controlling the shape and ownership of mutable messages-they must not have internal aliases and can only be owned by a single actor at a time. We demonstrate a static analysis built around isolation type qualifiers to enforce these constraints. Kilim comfortably scales to handle hundreds of thousands of actors and messages on modest hardware. It is fast as well-task-switching is 1000x faster than Java threads and 60x faster than other lightweight tasking frameworks, and message-passing is 3x faster than Erlang (currently the gold standard for concurrency-oriented programming).
Uploads
Papers by sriram srinivasan