CS 61C Spring 2010 TA: Michael Greenbaum
Section 114/118 Week 13 – Caches cs61c-tf@inst.berkeley.edu
notes
originally
by
Matt
Johnson
More Pipelining Topics!
Hazards: Our pipelined execution doesn’t always go smoothly. There are three
specific types of conflicts that can arise: what are they?
Superscalar Hardware: Some pipeline problems happen when we don’t have
enough hardware (e.g. “If only I had two dryers…”). Superscalar means adding
that redundant hardware for some instruction-level parallelism! Would that help
latency or throughput? Who decides what hardware is used when?
Out-of-Order Execution: After instruction fetch, dispatch to a queue where the
instruction waits until input operands are available. Queue is not always FIFO!
Results are queued too, and write-back order happens in the “original” instruction
order. Pretty complex! Low-end processors still do not use this paradigm due to
the silicon area that is required for its implementation… In this class, OoOE
would essentially refer to hardware doing the work of filling branch/load delay
slots. The lecture slides show an instruction pausing in the middle of its
execution.
Caches!
Conceptual Questions: Why do we cache? What is the end result of our
caching, in terms of capability?
What are temporal and spatial locality? Give high level examples in software of
when these occur.
Break up an address:
Tag
Index
Offset
Offset: “column index” (O bits)
Index: “row index” (i bits)
Tag: “cache number” that the block/row* came from. (T bits) [*difference?]
Segmenting the address into TIO implies a geometrical structure (and size)
on our cache. Draw memory with that same geometry!
Cache
Memory
2O
columns
Tag
=
0
i
2
Tag,
2i+O
Bytes
of
Valid,
&
rows
2T
Cache
Data!
Dirty
bits
Tag
=
1
Images
Tag
=
2
CS 61C Spring 2010 TA: Michael Greenbaum
Section 114/118 Week 13 – Caches cs61c-tf@inst.berkeley.edu
notes
originally
by
Matt
Johnson
Cache Vocab:
Cache hit – found the right thing in the cache! Booyah!
Cache miss – Nothing in the cache block we checked, so read from
memory and write to cache!
Cache miss, block replacement – We found a block, but it had the wrong
tag!
Cache Exercises!
C1: Fill this one in… Everything here is Direct-Mapped!
Address Cache Block Tag Bits Index Offset Bits per
Bits Size Size Bits Bits Row
16 4KB 4B
16 16KB 8B
32 8KB 8B
32 32KB 16B
32 64KB 16 12 4 146
32 512KB 5
64 64B 14
64 2048KB 1069
C2: Assume 16 B of memory and an 8B direct-mapped cache with 2-byte blocks.
Classify each of the following memory accesses as hit (H), miss (M), or miss with
replacement (R).
a. 4
b. 5
c. 2
d. 6
e. 1
f. 10
g. 7
h. 2
CS 61C Spring 2010 TA: Michael Greenbaum
Section 114/118 Week 13 – Caches cs61c-tf@inst.berkeley.edu
notes
originally
by
Matt
Johnson
C3: This composite question was inspired by exam questions but NOT identical
since the exam questions use associative caches. Direct-mapped here!
You know you have 1 MiB of memory (maxed out for processor address size)
and a 16 KiB cache (data size only, not counting extra bits) with 1 KiB blocks.
#define NUM_INTS 8192
int *A = malloc(NUM_INTS * sizeof(int)); // returns address 0x100000
int i, total = 0;
for (i = 0; i < NUM_INTS; i += 128) A[i] = i; // Line 1
for (i = 0; i < NUM_INTS; i += 128) total += A[i]; // Line 2
a) What is the T:I:O breakup for the cache (assuming byte addressing)?
b) Calculate the hit percentage for the cache for the line marked “Line 1”.
c) Calculate the hit percentage for the cache for the line marked “Line 2”.
d) How could you optimize the computation?
Now a completely different setup… Your cache now has 8-byte blocks and 128
rows, and memory has 22 bit addresses. The ARRAY_SIZE is 4 MiB and A starts at
a block boundary.
for (i = 0; i < (ARRAY_SIZE/STRETCH); i += 1) {
for (j = 0; j < STRETCH; j += 1) sum += A[i*STRETCH + j];
for (j = 0; j < STRETCH; j += 1) product *= A[i*STRETCH + j];
}
a) What is the T:I:O breakup for the cache (assuming byte addressing)?
b) What is the cache size (data only, no tag and extra bits) in bytes?
c) What is the largest STRETCH that minimizes cache misses?
d) Given the STRETCH size from (c), what is the # of cache misses?
e) Given the STRETCH size from (c), if A does not start at a block boundary,
roughly what is the # of cache misses for this case to the number you
calculated in question (d) above? (e.g., 8x, 1/16th)
