18-742 Test 1 October 8, 1997 Printed Initials: ________

18-742 Advanced Computer Architecture

Exam I

October 8, 1997

Name (please print): _____________________________________________________

Instructions: YOU HAVE 90 MINUTES TO COMPLETE THIS TEST

DO NOT OPEN TEST UNTIL TOLD TO START

The exam is composed of four problems containing a total of 14 sub-problems, adding up to 100
points overall. The point value is indicated for each subproblem and is roughly proportional to
difficulty. Attempt all problems and budget your time according to the problem’s difficulty. Show
all work in the space provided. If you have to make an assumption then state what it was. Answers
unaccompanied by supporting work will not be graded. The exam is closed book, closed notes, and
closed neighbors. You are on your honor to have erased any course-relevant material from your
calculator. If there is a single numeric answer to a problem, please circle it so that it stands out on
the page and we grade according to the answer you want to have considered as your best shot.
Please print your initials at the top of each page in case the pages of your test get accidentally
separated.

Good luck!

1
18-742 Test 1 October 8, 1997 Printed Initials: ________

Grading Sheet

1a) (7) ___________________

1b) (6) ___________________

2a) (10) ___________________

2b) (14) ___________________

3a) (6) ___________________

3b) (9) ___________________

3c) (7) ___________________

4a) (5) ___________________

4b) (6) ___________________

4c) (6) ___________________

4d) (6) ___________________

4e) (10) ___________________

4f) (8) ___________________

TOTAL: (100) _____________________

2
18-742 Test 1 October 8, 1997 Printed Initials: ________

1. Cache Policies.

Consider two alternate caches, each with 4 sectors holding 1 block per sector and one 32-bit word
per block. One cache is direct mapped and the other is fully associative with LRU replacement
policy. The machine is byte addressed on word boundaries and uses write allocation with write
back.

1a) (7 points) What would the overall miss ratio be for the following address stream on the direct
mapped cache? Assume the cache starts out completely invalidated.

read 0x00
read 0x04
write 0x08
read 0x10
read 0x08
write 0x00

1b) (6 points) Give an example address stream consisting of only reads that would result in a lower
miss ratio if fed to the direct mapped cache than if it were fed to the fully associative cache.

3
18-742 Test 1 October 8, 1997 Printed Initials: ________

2. Virtual Memory and Cache Organization.

The 742LX is a uniprocessor having up to a maximum of 64 MB of addressable physical

memory. The cache, virtual memory, and TLB have the following attributes:

Cache Virtual Memory TLB

unified virtual page size is 4 KB unified
physically addressed virtual address space is 1 GB fully associative
cache holds 20 KB 40 entries
5 way set associative 1 byte control/entry
32 Byte block size
sector size = block size
LRU replacement
write back
byte addresses on word boundaries

2a) (10 points) Sketch a block diagram of how the virtual address is mapped into the physical
address (assuming a TLB hit). Be sure to label exactly which/how many of the address bits go
where. and how many bits are in each of the 3 fields in a TLB entry.

4
18-742 Test 1 October 8, 1997 Printed Initials: ________

2b) (14 points) Given that you have the address output of a TLB and the original virtual address,
sketch a block diagram of how the cache is accessed to determine whether there is a cache hit (you
may ignore data access -- just indicate enough to say whether a hit or miss occurs; also include
only tag fields in your picture of the cache organization). Again label exactly which/how many
address bits go where and how big an address tag is.

5
18-742 Test 1 October 8, 1997 Printed Initials: ________

3. Multi-Level Caches.

You have a computer with two levels of cache memory and the following specifications:
CPU Clock: 200 MHz Bus speed: 50 MHz
Processor: 32-bit RISC scalar CPU, single data address maximum per instruction
L1 cache on-chip, 1 CPU cycle access
block size = 32 bytes, 1 block/sector, split I & D cache
each single-ported with one block available for access, non-blocking
L2 cache off-chip, 3 CPU cycles transport time (L1 miss penalty)
block size = 32 bytes, 1 block/sector, unified single-ported cache, blocking, non-pipelined
Main memory has 12+4+4+4 CPU cycles transport time for 32 bytes (L2 miss penalty)

Below are the results of a dinero simulation for the L1 cache:

CMDLINE: dinero -b32 -i8K -d8K -a1 -ww -An -W8 -B8
CACHE (bytes): blocksize=32, sub-blocksize=0, wordsize=8, Usize=0,
Dsize=8192, Isize=8192, bus-width=8.
POLICIES: assoc=1-way, replacement=l, fetch=d(1,0), write=w, allocate=n.
CTRL: debug=0, output=0, skipcount=0, maxcount=10000000, Q=0.

Metrics Access Type:

(totals,fraction) Total Instrn Data Read Write Misc
----------------- ------ ------ ------ ------ ------ ------
Demand Fetches 10000000 7362210 2637790 1870945 766845 0
1.0000 0.7362 0.2638 0.1871 0.0767 0.0000
Demand Misses 52206 8466 43740 36764 6976 0
0.0052 0.0011 0.0166 0.0196 0.0091 0.0000

Words From Memory 180920

( / Demand Fetches) 0.0181
Words Copied-Back 766845
( / Demand Writes) 1.0000
Total Traffic (words) 947765
( / Demand Fetches) 0.0948

3a) (6 points) What is the available (as opposed to used) sustained bandwidth between:

- L1 cache bandwidth available to CPU (assuming 0% L1 misses)?

- L2 cache bandwidth available to L1 cache (assuming 0% L2 misses)?

- Main memory bandwidth available to L2 cache?

6
18-742 Test 1 October 8, 1997 Printed Initials: ________

3b) (9 points) How long does an insruction take to execute (in ns), assuming 1 clock cycle per
instruction in the absence of memory hierarchy stalls, no write buffering at the L1 cache level, and
0% L2 miss rate.

3c) (7 points) A design study is performed to examine replacing the L2 cache with a victim cache.
Compute a measure of speed for each alternative and indicate which is the faster solution. Assume
the performance statistics are
L2 cache local miss ratio = 0.19
Victim cache miss ratio = 0.26; and its transport time from L1 miss = 1 clock

7
18-742 Test 1 October 8, 1997 Printed Initials: ________

4. Address Tracing and Cache Simulation.

You have instrumented only data references from the subroutine “sum_array” in the following
program using Atom on an Alpha workstation (“long” values are 64 bits). The resultant data reads
and writes have been run through dinero with a particular cache configuration. In this question
you’ll deduce the cache configuration used.

#include <stdio.h>
#include <stdlib.h>

void sum_array(int N, long *a, long *b, long *c, long *d, long *e)
{ int i;
for (i = 0; i < N; i++)
{ *(a++) = *(b++) + *(c++) + *(d++) + *(e++); }
}

int main(int argc, char *argv[])

{ int N, offset, i;
long *a, *b, *c, *d, *e; /* 64-bit elements in each array */

if (argc != 3)
{ fprintf(stderr, "\nUsage: test <size> <offset>\n"); exit(-1); }
sscanf(argv[1], "%d", &N); sscanf(argv[2], "%d", &offset);

a = (long *) malloc(5 * ( (N*sizeof(long)) + offset) );

b = a + N + ( offset/sizeof(long) );
c = b + N + ( offset/sizeof(long) );
d = c + N + ( offset/sizeof(long) );
e = d + N + ( offset/sizeof(long) );

sum_array(N, a, b, c, d, e);

The program was executed with a command line having successively higher values of N from 1 to
2100, and an offset value of 0. The below graph shows the number of combined data misses for
each value of N.
• Bus size and word size are both 8 bytes.
• The cache has one block per sector and a block size of 128 bytes (16 words).
• Assume a completely invalidated cache upon entry to sum_array.

8
18-742 Test 1 October 8, 1997 Printed Initials: ________

Total Data Cache Misses

6500

6000
5500

5000
4500

4000
3500
Series1
3000

2500

2000
1500

1000
500
0
0 500 1000 1500 2000 2500
Value of N

Some data that might be of interest are:

N # Misses Total Traffic (words)
500 626 2516
501 628 2533
502 629 2534
503 630 2535
504 630 2520
505 632 2537
506 633 2538
507 696 3531
508 1020 8700
509 1277 12797
510 1597 17902
511 1853 21983
512 2079 25584
513 1862 22097
514 1606 17986
515 1288 12883
516 1032 8772
517 717 3717
518 651 2646
519 653 2663
520 652 2632
521 655 2665
522 655 2650
523 657 2667
524 658 2668

Answer the following questions, giving brief support for your answer of (unsupported answers
will not receive full credit).

9
18-742 Test 1 October 8, 1997 Printed Initials: ________

4a) (5 points) Ignoring overhead for the subroutine call, what is the theoretical minimum possible
data total traffic (in 8-byte words) that this program has to move (combined into and out of the
cache) for N = 500?

4b) (6 points) Does the cache perform write allocation?

4c) (6 points) Assuming it is not direct mapped, does this data look like it came from an LRU
replacement policy or a random replacement policy? Why?

10
18-742 Test 1 October 8, 1997 Printed Initials: ________

4d) (6 points) Assume that you have a direct mapped cache. What is the best value for the input
parameter offset if you want to improve performance for N=512 (“best” means the smallest
value guaranteed to have 100% effectiveness for N=512).

4e) (10 points) What is the actual associativity of the cache that produced the data given? (and
how did you figure that out?)

4f) (8 points) How many bytes does the cache hold (data only, not counting control+tag bits)?

11
18-742 Test 1 February 24, 1998 Printed Initials: ________

18-742 Advanced Computer Architecture

Test I

February 24, 1998

Name (please print): _____________________________________________________

Instructions: DO NOT OPEN TEST UNTIL TOLD TO START

YOU HAVE UNTIL 12:20 PM TO COMPLETE THIS TEST

The exam is composed of four problems containing a total of 11 sub-problems, adding up to 100
points overall. The point value is indicated for each sub-problem and is roughly proportional to
difficulty. Attempt all problems and budget your time according to the problem’s difficulty. Show
all work in the space provided. If you have to make an assumption then state what it was. Answers
unaccompanied by supporting work will not receive full credit. The exam is closed book, closed
notes, and “closed neighbors.” You are on your honor to have erased any course-relevant material
from your calculator prior to the start of the test. Please print your initials at the top of each page
in case the pages of your test get accidentally separated. You may separate the pages of the test if
you like, and re-staple them when handing the test in.

Good luck!

1
18-742 Test 1 February 24, 1998 Printed Initials: ________

Grading Sheet

1a) (5) ___________________

1b) (8) ___________________

2a) (9) ___________________

2b) (5) ___________________

2c) (10) ___________________

2d) (10) ___________________

3a) (9) ___________________

3b) (10) ___________________

4a) (9) ___________________

4b) (10) ___________________

4c) (15) ___________________

TOTAL: (100) _____________________

2
18-742 Test 1 February 24, 1998 Printed Initials: ________

1. Cache Policies.

Consider a cache memory with the following characteristics:

• Unified, direct mapped
• Capacity of 256 bytes (0x100 bytes, where the prefix “0x” denotes a hexadecimal number)
• Two blocks per sector
• Two words per block
• 4-byte (32-bit) words, addressed in bytes but accessed only in 4-byte word boundaries
• Bus transfers one 32-bit word at a time, and transfers only those words actually required to be
read/written given the design of the cache (i.e., it generates as little bus traffic as it can)
• No prefetching (i.e., demand fetches only; one block allocated on a miss)
• Write back policy; Write allocate policy

1a) (5 points) Assume the cache starts out completely invalidated. Circle one of four choices next
to every access to indicate whether the access is a hit, or which type of miss it is. Put a couple
words under each printed line giving a reason for your choice.

read 0x0010 hit, compulsory miss, capacity miss, conflict miss

read 0x0014 hit, compulsory miss, capacity miss, conflict miss

read 0x001C hit, compulsory miss, capacity miss, conflict miss

read 0x0110 hit, compulsory miss, capacity miss, conflict miss

read 0x001C hit, compulsory miss, capacity miss, conflict miss

1b) (8 points) Assume the cache starts out completely invalidated. Circle “hit” or “miss” next to
each access, and indicate the number of words of memory traffic generated by the access.

write 0x0024 hit, miss, # words bus traffic = _________

write 0x0020 hit, miss, # words bus traffic = _________

write 0x0124 hit, miss, # words bus traffic = _________

read 0x0024 hit, miss, # words bus traffic = _________

3
18-742 Test 1 February 24, 1998 Printed Initials: ________

2. Multi-level cache organization and performance.

Consider a system with a two-level cache having the following characteristics. The system does
not use “shared” bits in its caches.

L1 cache L2 cache
- Physically addressed, byte addressed - Physically addressed, byte addressed
- Split I/D - Unified
- 8 KB combined, evenly split -- 4KB each - 160 KB
- Direct mapped - 5-way set associative; LRU replacement
- 2 blocks per sector - 2 blocks per sector
- 1 word per block (8 bytes/word) - 1 word per block (8 bytes/word)
- Write-through - Write-back
- No write allocation - Write allocate
- L1 hit time is 1 clock - L2 hit time is 5 clocks (after L1 miss)
- L1 average local miss rate is 0.15 - L2 average local miss rate is 0.05

- The system has a 40-bit physical address space and a 52-bit virtual address space.
- L2 miss (transport time) takes 50 clock cycles
- The system uses a sequential forward access model (the “usual” one from class)

2a) (9 points)
Compute the following elements that would be necessary to determine how to configure the cache
memory array:

Total number of bits within each L1 cache block: __________________

Total number of bits within each L1 cache sector: __________________

Total number of bits within each L2 set: __________________

4
18-742 Test 1 February 24, 1998 Printed Initials: ________

2b) (5 points)
Compute the average effective memory access time (tea, in clocks) for the given 2-level cache under
the stated conditions.

2c) (10 points)

Consider a case in which a task is interrupted, causing both caches described above to be
completely flushed by other tasks, and then that original task begins execution again (i.e., a
completely cold cache situation) and runs to completion, completely refilling the cache as it runs.
What is the approximate time penalty, in clocks, associated with refilling the caches when the
original program resumes execution? A restating of this same question is: assuming that the
program runs to completion after it is restarted, how much longer (in clocks) will it take to run
than if it had not been interrupted? For this sub-problem:
• The program makes only single-word memory accesses
• The reason we say “approximate” is that you should compute L1 and L2 penalties
independently and then add them, rather than try to figure out coupling between them.
• State any assumptions you feel are necessary to solving this problem.

5
18-742 Test 1 February 24, 1998 Printed Initials: ________

2d) (10 points)

For the given 2 level cache design, sketch the data cache behavior of the following program in the
graphing area provided (ignore all instruction accesses -- consider data accesses only). Label each
point of inflection (i.e., points at which the graph substantially changes behavior) and the end
points on the curve with a letter and put values for that point in the data table provided along with
a brief statement (3-10 words) of what is causing the change in behavior and how you knew what
array size it would change at. The local miss rates given earlier are for a “typical” program, and
obviously do not apply to the program under consideration in this sub-question.

long touch_array(long *array, int size)

{ int i;
long sum=0;
long *p, *limit;
limit = array + size/sizeof(int);
for (i = 0; i < 1000000; i++)
{ /* sequentially touch all elements in array */
for (p = array; p != limit; p++)
{ sum += *p; }
}
return(sum);
}

Point Array size tea Reason

A 1 KB 1 All L1 Hits
B
C
D
E
F 1000 KB All L2 Misses

6
18-742 Test 1 February 24, 1998 Printed Initials: ________

3. Virtual Memory.

Consider a virtual memory system with the following characteristics:

• Page size of 16 KB
• Page table and page directory entries of 8 bytes per entry
• Inverted page table entries of 16 bytes per entry
• 31 bit physical address, byte addressed
• Two-level page table structure (containing a page directory and one layer of page tables)

3a) (9 points) What is size (in number of address bits) of the virtual address space supported by
the above virtual memory configuration?

3b) (10 points) What is the maximum required size (in KB, MB, or GB) of an inverted page table
for the above virtual memory configuration?

7
18-742 Test 1 February 24, 1998 Printed Initials: ________

4. Cache simulation

The following program was instrumented with Atom and simulated with a cache simulator. The
cache simulator was fed only the data accesses (not instruction accesses) for the following
subroutine:

long test_routine(long *a, int size, int stride)

{ long result=0;
long *b, *limit;
int i;

limit = a + size;
for (i = 0; i < 100; i++)
{ b = a;
while (b < limit)
{ result += *b;
b += stride;
}
}
return(result);
}

The result of sending the Atom outputs through the cache simulator for various values of “size”
and “stride” (each for a single execution of test_routine) are depicted in the graph below.
The cache word size was 8 bytes (matching the size of a “long” on an Alpha), and a direct-mapped
cache was simulated. The simulated cache was invalidated just before entering test_routine.
There was no explicit prefetching used. The code was compiled with an optimization switch that
got rid of essentially all reads and writes for loop handling overhead.
Use only “Cragon” terminology for this problem (not dinero terminology).

0.9

0.8 Value of
"size"
0.7
256
Miss Ratio

0.6 512
1024
0.5
2048
0.4 4096
8192
0.3

0.2

0.1

0
0 4 8 12 16 20 24 28 32

Value of "stride"

8
18-742 Test 1 February 24, 1998 Printed Initials: ________

4a) (9 points)
For the given program and given results graph, what was the size of the cache in KB (support your
answer with a brief reason)? Remember that in the C programming language pointer arithmetic is
automatically scaled according to the size of the data value being pointed to (so, in this code,
adding 1 to a pointer actually adds the value 8 to the address).

4b) (10 points)

For this example, what is the cache block size in bytes (support your answer with a brief reason)?

9
18-742 Test 1 February 24, 1998 Printed Initials: ________

4c) (15 points)

What parameter is responsible for the “size”=2K (the data plotted with circles) point of inflection
at a “stride” of 16? What is the value of that parameter, and why?

10