Simulating computation

Simulating computation

While the statements described in Delaying execution delay for a specified length of time, it is also possible to delay for the duration of a specified amount of “work”. “Work” is expressed in terms of memory accesses. That is, a coNCePTuaL program can touch (i.e., read plus write) data with a given stride from a memory region of a given size. By varying these parameters, a program can emulate an application’s computation by hoarding the CPU or any level of the memory hierarchy.

<item_size> and <data_type> are described in Item size and <expr> is described in Arithmetic expressions.

As shown by the formal definition of <touch_stmt> the required components are a <source_task> and the size of the memory region to touch. By default, every WORD (see Item size) of memory in the region is touched exactly once. The type of data that is touched can be varied with an ‘ <expr> <data_type> OF’ clause. For instance, ‘100 BYTES OF’ of a memory region will touch individual bytes. An optional repeat count enables the memory region (or subset thereof) to be touched multiple times. Hence, if ‘TASK 0 TOUCHES A 6 MEGABYTE MEMORY REGION 5 TIMES’, then the touch will be performed as if ‘TASK 0’ were told to ‘TOUCH 5*6M BYTES OF A 6 MEGABYTE MEMORY REGION 1 TIME’ or simply to ‘TOUCH 5*6M BYTES OF A 6 MEGABYTE MEMORY REGION’.

By default, every <data_type> of data is touched. However, a <touch_stmt> provides for touching only a subset of the <data_type>s in the memory region. By writing ‘WITH STRIDE <expr> <data_type>’, only the first <data_type> out of every <expr> will be touched. Instead of specifying an exact stride, the memory region can be accessed in random order using the WITH RANDOM STRIDE clause.

Unless the number of touches and data type are specified explicitly, the number of WORDs that are touched is equal to the size of the memory region divided by the stride length then multiplied by the repeat count. Therefore, if ‘TASK 0 TOUCHES AN 8 MEGABYTE MEMORY REGION 2 TIMES WITH STRIDE 8 WORDS’, then a total of (2^23 / (4*8)) * 2 = 524288 touches will be performed. For the purpose of the preceding calculation, ‘WITH RANDOM STRIDE’ should be treated as if it were ‘WITH STRIDE 1 WORD’ (again, unless the number of touches and data type are specified explicitly).

To save memory, all TOUCH statements in a coNCePTuaL program access subsets of the same region of memory, whose size is determined by the maximum needed. However, each dynamic execution of a <touch_stmt> starts touching from where the previous execution left off. For example, consider the following statement:

TASK 0 TOUCHES 100 WORDS OF A 200 WORD MEMORY REGION

The first time that that statement is executed within a loop (see Iterating), the first 200 words are touched. The second time, the second 200 words are touched. The third time, the index into the region wraps around and the first 200 words are touched again.

Each static <touch_stmt> maintains its own index into the memory region. Therefore, the first of the following two statements will terminate successfully (assuming it’s not executed in the body of a loop) while the second will result in a run-time error because the final byte of the final word does not fit within the given memory region.

TASK 0 TOUCHES 100 WORDS OF A 799 BYTE MEMORY REGION THEN
TASK 0 TOUCHES 100 WORDS OF A 799 BYTE MEMORY REGION

FOR 2 REPETITIONS TASK 0 TOUCHES 100 WORDS OF A 799 BYTE MEMORY REGION

( THEN is described in Combining statements, and FOR… REPETITIONS is described in Counted loops.) The first statement shown above touches the same 100 words (400 bytes) in each of the two <touch_stmt>s. The second statement touches the first 100 words the first time the <touch_stmt> is executed and fails when trying to touch the (only partially extant) second 100 words.