Abdellah Touhafi

SURReCA

Home | Contact | Research Interests | Education | Rec. ComputingAcoustics | Thesis Subjects |Publications

SURReCA

Download PDF
will be available soon!

SURReCA : a scalable runtime

reconfigurable computing system
 

SURReCA (Scalable Uniform Runtime Reconfigurable Computing Architecture) is a scalable runtime reconfigurable computer that consists of a set of distributed computing nodes, which can be interconnected in a mesh structure. The computing nodes can be runtime reconfigured which means that the structure of their arithmetic and logic unit can be chaned at runtime. This in order to change the computing structure in respect to the needs of the application. A prototype system that contains four computing nodes has been built. See
Figure 1.

The SURReCA node architecture consists of three parts: a part that contains the computing elements, a part that contains the memory resources and a centrally placed part that implements the reconfiguration engine. The computing layer and the configuration engine have both four communication-links, such that the nodes can be interconnected in a mesh structure. This way, the RTR system can benefit from a large amount of computing resources, with a tightly interconnected reconfiguration engine of which the reconfiguration bandwidth scales linearly with the amount of computing resources.

The computing layer contains the computing resources, the local data memory and two buffer controllers. One controller keeps track of the output buffer state and one controller is used to check the state of the input buffer. The input buffer is to be interpreted as a source buffer of a context and the output buffer implements a sink buffer in which data values are stored.

Figure 1: SURReCA prototype

 

 

Control

Layer

The control layer contains the configuration engine that is the base of the distributed controller. Important is that the configuration engine is implemented as a programmable unit that uses the structure information elements together with some instructions to control the computing layer and to synchronize the distributed system. All the information elements of the contexts are grouped into a fixed record structure and are located in the contexts memory. The SURReCA node contains also an instructions memory that is used to specify the reconfiguration function. For clarity a separated bus for the instruction memory, the context memory and the application data memory is drew. In practical realization the busses are shared because of reasons explained in the previous chapter. Multi-node SURReCA systems can be built by interconnecting the communication links of the nodes in a mesh structure or another convenient structure.

Instruction

set

 

The DCC computing model represents a computing problem by a reconfiguration graph and a structures graph. In that optic, the handling of the reconfiguration graph is associated with the reconfiguration controller. The reconfiguration controller implements the reconfiguration function and the state transition function of the DCC modelís reconfiguration manager. For the specification of the state transition function and the reconfiguration function, an instruction set is defined. An application is hence specified as a program containing instructions, instruction parameters, context pointers and pointers to buffers.The controller determines what actions must be taken to change a structure by means of the instructions. For the implementation of a complete DCC specified application, only three instructions are needed to control the whole distributed system.

-JMP
The first instruction is an absolute conditional jump instruction that is used to specify the flow of the configuration graph. The jump instruction has two parameters. The first parameter contains the jump condition that must be checked and the second parameter contains the jump address that must be jumped to. The jump address is an absolute address.

JMP

Jmp_condition

Jmp_address

 -SYNCH
The second instruction is a synchronization instruction that is used to specify with which neighbors, an activation must be synchronized. The instruction has two parameters. One parameter to specify the neighbors that must be synchronized with and a synchronization-tag that is used to discriminate among different synchronizations. All nodes that must be synchronized will use an equal Synch_tag.
 

SYNCH

Synch_tag

Neighbors

 The synchronization goes in two phases. When the Synch instruction is started the synchronization register will contain the neighbors that must be synchronized with and the synch_tag register will contain the synchronization TAG. When a node is configured, it will send the synch_tag to some of its neighbors to make clear that it is ready to start computing. The node will then wait for a global synch signal that is used to synchronize the computing nodes. The synchronization controller accomplishes the whole synchronization procedure and uses the link controller as a communication medium to send and receive synchronization tags. The synchronization procedure is detailed later in the text.

 -MOVE
The last instruction is a 'structure move' instruction that is used to specify which contexts must be activated. The format of that instruction is:

MOVE

Pointer to Structure Information Record

Stop_condition

The instruction has also two parameters, the first parameter specifies a pointer to a record that contains all the information about the structure that must be activated and the second parameter specifies the deactivation condition. The idea is that once a structure is activated it will be kept active until the specified stop_condition is met. In case the next instruction is also a move instruction, the handling of that instruction will be blocked until the active context is deactivated. In case the next instruction is a SYNCH instruction, that instruction will be handled during the configuration phase. In such a case the SYNCH-instruction acts as a start condition for the configured context.

The stop condition is defined as four two bit values, where each value is associated with a context switcher output. The values in the register test for the following context switcher output:

Value

CS output

00

False

01

True

10

Donít Care

11

Donít Care

The information of the structural setup is stored in a structure information record (SIR) to which the SIR-register points The organization of the record is as shown here.
The record starts with specifying the number of contexts, which the structure contains. Then for all the contexts there is a context ID, a context pointer that points to the configuration, which must be loaded followed by the state information of that context and a delimiter. After the delimiter for the last context, the record contains the interconnect setup information followed by a delimiter for the structure.

Number of contexts

Context ID

Context pointer

State information

Ö

End_delimiter for first context

Context ID

Ö

End_delimiter for last context

Interconnect setup

End_delimiter Structure