Project
A
Optimization
of SPIDER for Highly Parallel Machines
Director:
Joachim Frank, Wadsworth Center
The aim of this project is to adapt SPIDER to run on massively parallel
processor platforms such as Linux Beowulf clusters or a group of
distributed, networked workstation running under the Unix or NT
operating system. SPIDER (System
for Processing Image Data in Electron microscopy and Related fields) is currently
parallelized only on a fine-grained level of computing, using the Open
Multiprocessing directives for Shared memory Processing that are placed
within the SPIDER Fortran code. We intend to add high-level
coarse-grain parallelism at the process or SPIDER job level, in a way
that does not require shared memory. Processing will be
partitioned by groups of images such as defocus groups, or on the level
of single images.
The way this will be implemented is through a system of scheduling
independent jobs on different processors. Starting with an
initial serial or parallel SPIDER job, the availability of jobs will be
"published" to all processors. Each available processor runs a
small active subscription process that looks at the queue for pending
jobs, "subscribes" to one, and locks it to make it unavailable to other
subscriptions. Upon normal termination, the locked job is also
removed from the shared queue. Abnormal termination will free the
job up for another subscription.
We intend to augment this approach in collaboration with Project C by
inserting powerful Message Passing Interface (MPI) subroutine calls
into our code, which will take advantage of homogeneous clusters
wherever these are available. Together, we will develop and test
implementations of SPIDER code that make use of both Publish &
Subscribe and MPI. Specific computationally intensive parts of
the code, such as orientation search and alignment (AP_MQ in the SPIDER
suite of commands) and iterative algebraic reconstruction, will be
analyzed for opportunities of parallelization with MPI under the Single
Program Multiple Data programming model.
This overall plan entails the need for the development of specific
tools within SPIDER for error recovery, check pointing, and message
passing, all of which will be addressed in this proposal. For the
final design of the system we propose to have master publish and
subscribe processes written in Python.
Another aim is to also develop a high-level organizational framework
for image processing and reconstruction, in a way that facilitates
overall optimization and backtracking based on complete records of
data. While the first aim addresses numerical-processing speed
without regard to the contents of the operations, this aim recognizes
the lack of an overall intelligence in the execution of a series of
algorithms where each has been optimized with attention to local, not
global needs.
While this second aim marks a radical departure from the philosophy
with which SPIDER was designed (with maximum flexibility on the lowest
level) we will refrain from any radical redesign of the software, as a
matter of practicality and to save an enormous investment of many
programmer years. Rather, our approach will be to keep the SPIDER
command structure and syntax in place, and design a higher-level
"shell" structure for the control of the processing following current
SPIDER procedures, as well as for the design of new SPIDER
procedures.
Project A will interact extensively with Project B. The above
mentioned aims will be done in liaison with Project B. there are
many aspects of parallelization, algorithm design, and reconstruction
strategies that impact both software implementations in similar
ways. We will jointly decide how to break down the entire
reconstruction procedure into large interchangeable modules with
standardized interfaces. The purpose of this plan is threefold:
(1) make benefits of either package, at every step of development,
available to all users; (2) facilitate comparative performance testing;
and (3) enforce standardization of parameters, options, and
procedures.
.
