next up previous contents index
Next: Full Matrix Algorithms Up: Parallel Computing in Previous: 7.6.5 Task Farming

7.6.6 Distributed Modelling via the Postmaster Element

The postmaster element is a self-contained object within the GENESIS neural simulator. Like the objects in a true object-oriented language it is an entity composed of public (externally available) data, private (internal) data, and functions to operate on the object. Like other GENESIS elements, it is a black box that can be used to construct a neural simulation. For the present, modellers wishing to produce large distributed simulations that are able to take full advantage of the power inherent in our current parallel computers must specify how to distribute the different parts of their simulation over the available hardware computing nodes. While this may be an annoying necessity to certain neural modellers, it must be remembered that this technique allows the construction of simulations of such a size and computational complexity that they can be modelled on no other existing computer platforms (e.g., traditional serial supercomputers). Bearing this in mind, the requirement of explicitly stating how to distribute the model is a small price to pay for the power gained. This explicit method also brings with it certain advantages. Firstly, because the modeller is familiar with the computational demands of the different parts of his model, he is able to accurately balance the computational load over the available parallel hardware. This makes for far more efficient load balancing and scaling behavior than would be possible with an automatic decomposition scheme that has to balance the very different needs of both single-cell modellers and network-scale modellers. It is also less efficient to carry out automatic decomposition because the various GENESIS elements comprising a complete neural simulation have widely varying computational demands. As with other GENESIS elements, the postmaster acts as a self-contained black box which usually communicates information about its changing state via messages to other elements. It can both send messages to, and receive messages from, other GENESIS elements which may exist either on this particular hardware node or on others in the network. As a result of this, it is an element that ties together and coordinates the disparate parts of a model running on separate hardware nodes of the parallel computer. The actual messages transferred depend on the type of element with which it communicates. The postmaster element neither knows nor cares whether the quantity it is communicating is a membrane potential, a channel conductance (simple current flow), or the concentration of any substance from ions to complex neurotransmitters. The postmaster is a two-faced element. Its first face is that of a normal GENESIS element which it presents to the rest of the simulation. This first aspect of the postmaster is able to pass GENESIS messages between elements and allows use of the GENESIS Script Language to query its state. The second aspect of the postmaster is aware of the parallel nature of the underlying hardware and can use the operating system primitives for communicating information between separate physical nodes. In summary, the postmaster element is a conduit along which information can flow between nodes. It allows the modeller to tie together the disparate aspects of the simulation into a coherent whole.

To show how this mechanism works in practice, performance measurements are presented in Figure 7.14, which are taken on the Intel Touchstone Delta parallel supercomputer based at Caltech. These figures show both scaling of performance, where more nodes allow the model to complete in a shorter timescale, and the construction of models so large that it has been impossible to model them before now.

  
Figure 7.14: Results from Running a GENESIS Simulation of a Passive Cable Model Composed of Varying Numbers of Compartments Distributed Across Varying Numbers of Nodes on the Intel Touchstone Delta Parallel Supercomputer

The previous record for the most complex GENESIS model produced on a traditional serial computer is approximately 80,000 elements, the Purkinje Cell  model produced at Caltech by Dr. de Schutter [Schutter:91a;93a]. Using the postmaster element on a parallel supercomputer (the 512-node Intel Touchstone Delta), this limit has now been pushed to over two million GENESIS elements (actually ). As can be seen, this now allows for the construction of far more complex and realistic models than was previously possible. The present price to be paid for this freedom is the decision to make the modeller explicitly distribute his simulation over the available hardware. This was a design decision that has allowed far greater efficiency of load balancing than would be possible using an automatic distribution technique as the illustrated scaling graphs confirm. This technique also has the advantage of leaving the basic GENESIS script interface unaltered and is applicable to a wide variety of parallel hardware. As a result of the requirement to retain compatibility with the existing serial GENESIS implementation, another benefit has become obvious. By changing only the network layer of the postmaster element, it is possible to produce a version of the postmaster that can use traditional serial machines distributed across the Internet to produce a distributed model, which ties together existing supercomputers based anywhere on the network. The potential size of model that can be constructed in this manner is staggering, although the reality of network communication delays will limit its area of application to compute-limited tasks (cf. communication-bound simulations).

The assessment of the usefulness of a parallel neural modelling platform can be demonstrated by two extremes of neural modelling applications:

  1. Detailed single-cell models

    The individual subcompartments which make up a neuron's dendritic tree are active simultaneously, and each of these compartments may be studded with several independently functioning ionic membrane channels. This is illustrated in Figure 7.15.


    Figure 7.15: Cerebella Purkinle cell model in GENESIS. Our most detailed single-cell model to date.

  2. Large network simulations

    A detailed network simulation may be composed of many thousands of individual nerve cells from a smaller number of biological cell types.

What does a parallel computer have to offer the single-cell modeller?

Most of the work on the parallel GENESIS at Caltech to date has been in the field of single-cell modelling. Several distinct ways of using the system have been developed allowing a variety of approaches by the single-cell modeller.

What does a parallel computer have to offer the neural network  modeller?

Much of the work to date has been on task farming (as described above), whereby each node runs its own copy of a cell. This is less useful to the network modeller but still allows detailed statistical information to be built up about network and population behavior. A more interesting way of using the parallel machine for network modelling is the distributed model scheme mentioned above. This allows networks to be both larger and to run more quickly than their counterparts on equivalent serial machines. A promising project in this category, although still in its very early stages, is the construction by Upinder Bhalla at Caltech of a detailed model of the rat olfactory bulb [Bhalla:93a]. This incorporates detailed cellular  elements, which are rare in network class models. Such a network model makes far greater communication demands of the internode communications mechanism on the parallel computer than a distributed single-cell model. Initially an expanded version of the postmaster element [Speight:92a;92b] which was used for distributed single-cell models will be employed, but this may change as the different demands of a large network model become apparent.

All of the work on the Parallel GENESIS project was carried out in the laboratory of Professor James Bower at the California Institute of Technology.



next up previous contents index
Next: Full Matrix Algorithms Up: Parallel Computing in Previous: 7.6.5 Task Farming



Guy Robinson
Wed Mar 1 10:19:35 EST 1995