Scintilla Overview

Cluster Equipment:

The Parallel Systems Lab at UCSB recently acquired a cluster of 8 ULTRASparc workstations (6 Ultra-1 and two Ultra-2 SMP's). These stations are connected by a high-speed, low-latency SCI (Scalable Coherent Interface) network provided by Dolphin Interconnect Solutions. We also experiment on a small cluster of four PCI-based Pentium Pro servers running Windows NT.  We will expand our Sparc-based cluster with more Ultra-1's and four PCI-based Ultra-30's. The combined cluster will be extremely heterogeneous, allowing us to experiment with a  real life workstation cluster one could expect in a typical industrial setting.

SCI Highlights:

SCI delivers a high bandwidth and a low latency for remote access (about 4us). SCI provides a computer-bus-like service, but uses point-to-point links instead of a physical bus to achieve high throughput. Therefore, SCI can be used for tightly coupled systems as well as networks of workstations (NOW). SCI supports efficient multiprocessor lock transactions, cache coherence in a shared-distributed-memory model, non-coherent caching, and message passing. All these features are standardized by ANSI/IEEE Std 1596-1992 SCI. Several companies and universities contributed to the definition of SCI, e.g. Apple, CERN, Dolphin, Hewlett Packard, University of Oslo, University of Wisconsin, and more.

Active Messages and Split-C:

On top of SCI we have implemented Active Messages for the communication between the nodes. Active Messages is an asynchronous communication mechanism intended to expose the full hardware flexibility and performance of modern interconnection networks. The underlying idea is simple: each message contains the address of a user-level handler which is executed upon message arrival. Handlers must execute quickly and complete without suspending. For example, a handler might write the message data directly into user space without using a deeply layered networking protocol. Active Messages provide an interface for higher level applications to provide easy to use, reliable and fast communication mechanisms.

 Split-C is an extension of the C programming language offering a global address space. It assumes a single program multiple data (SPMD) model in which each of the CPU's has a single thread of control and the memory model is a two-dimensional array, where one dimension is the set of CPU's, the other dimension is each processors local address space. Accesses to memory locations on a remote node are compiled to code fetching or putting data from or to that remote processor. Split-C allows to overlap communication of communication and computation by using split phase operations (called gets, puts and stores). Split-C is available on a variety of supercomputers, including the TMC CM-5, the Meiko CS-2, Intel Paragon and IBM SP-2 machines, as well as for networks of workstations. Most implementations, including one of our SCI implementations of Split-C, are based on Active Messages. We also have a second implementation of Split-C that directly uses the shared memory features of SCI to access remote memory, much like the Cray T3D implementation.

Fast Java Sockets and Parallel Web Servers:

We are currently investigating fast communication for Java. In particular, we are analyzing what the bottlenecks for native methods in SUN's JDK are, and how to overcome the obstacles encountered. We also work on a toy implementation of a parallel web server and study how shared-memory-based networks without efficient synchronization primitives like SCI  can be used to implement fast distributed caches.

Future Work:

Currently we run our first prototype of Active Messages and Split-C on our cluster consisting of four UltraSparc workstations and one Dolphin 4-way switch. We have presented preliminery performance measurements at the sixth and seventh SCIzzL workshops. We also implemented a software TLB (SWTLB) for SCI that allows for a more efficient implementation of  Split-C and presented a prototype in Hot-Interconnects 1997.

We are working on a more efficient SWTLB version of Split-C and start working on the issue of hybrid SMP-Clusters (CLUMPS) and non-dedicated computing, in particular co-scheduling.