Introduction

The Translation Lookaside Bu er, TLB, is a cache of page table entries that are used in virtual to physical address translation. Since the cache is of nite size we should be able to detect a di erence in the time it takes to perform a memory operation depending on if we nd the page table entry in the cache or not.

We will not have access to any operations that will explicitly control the TLB but we can access pages in a way that forces the CPU to have more or less page hits. We will set up a number of virtual pages in memory and if we access them linearly we should be able to see some changes when we have more pages than the TLB can handle.

1 Measuring time

To explore the di erence between hitting the TLB and having to nd the page table entry in memory, we will write a small benchmark program that accesses more and more pages to see when the number is to large to t in the TLB. In order to do this we rst of all need a way to measure time.

1.1 process or wall time

First of all we need to decide if we want to measure the process time or the wall time. The process time is the time our process is scheduled for execution whereas the wall time also includes the time the process is suspended. The process is suspended if the operating system has to schedule another process or if our process is waiting for something, for example I/O. In this experiment we will do ne with the process time.

Look up the manual entry for the procedure clock(), it will give us a time stamp that we can use to calculate the elapsed process time. There is then a macro, CLOCKS_PER_SEC, that we use to translate the timestamp into seconds. This is the time we would normally use if we are not doing any I/O or other things that we also want to include in our measurements.

1.2 the test rig

We want to measure how long time it takes to access a page and see if the time increases when we increase the number of pages. We will set a maximum number of pages (we will start with 16) and then measure the time it takes to access one page, two pages, three pages etc. To get more reliable gures we need to do a multiple of references in each test.In order to make it easier to compare the numbers we will make the same number of references in each test. We therefor set up a loop in a loop where the inner loop will access the number of pages in the test and the outer loop is adjusted to make the total number or references the same.

The increment of the sum is only there to have something to measure. We will later remove it and replace it what we really want to measure.

#include <stdlib .h>

#include <stdio .h> #include <time .h>

#define PAGES (16)

#define REFS (10241024) int main( int argc , char argv [ ] ) { clock_t c_start , c_stop ; printf (#pagest proct sum
); for ( int pages = 1; pages <= PAGES; pages+=1) { int loops = REFS / pages ; c_start = clock ( ) ; long sum = 0;

for ( int l = 0; l < loops ; l++) {

for ( int p = 0; p < pages ; p++) { / dummy operation / sum++;

}

}

c_stop = clock ( ) ;

{

double proc ; proc = (( double )( c_stop c_start ))/CLOCKS_PER_SEC; printf (%dt %.6 f t %ld
, pages , proc , sum );

}

}

return 0;

}

In the example above Ive used a million references but this might di er from machine to machine. Experiment with less or more references to see how many you would need to get measurements that do not uctuate to much. In the printf() statement you can change the sequence %.6ft to write more or less decimals. How many signi cant gures is it reasonable to show?

1.3 dont let i go

Do the numbers look right? Is the time to do the rst and second test slightly longer than the other tests? Is this just a glitch in the system or is it real? Hmm, there is something going on here? What is happening?

When you see strange numbers try to gure out what the problem could be. In this case the numbers show that it takes more time to run the outer loop one million times, where we inside the loop only loops once, compared to making the outer loop 500.000 times and the inner loop twice. If you think about it is what we would expect since there is an overhead in preparing for the inner loop and jumping out of the loop. The more time we spend inside the loop the less the overhead will make a di erence.

1.4 optimize but not too much

We will in the end try to estimate what it costs to do one memory reference. The problem is that test rig we now have set up actually makes tons of memory references by itself. The sum, l and p are probably stored on the stack so we will make multiple references to the stack in each test. We could try to reduce the memory overhead by turning on some optimizations. Try the following if you are using gcc.

> gcc -o tlb -O tlb.c

This will turn on the rst level of optimizations. I think that you have a substantial improvement. You can verify what happened by looking at the generated assembly. Using the -S ag we can ask the compiler to generate an assembler le, try the following:

> gcc -o tlb-opt.s -S -O tlb.c

:

> gcc -o tlb-reg.s -S tlb.c

Now compare the two les and see how they di er. Look for the call to clock where you will nd the loop. My guess is that the regular version will be lled with references like -16(%rbp) i.e. a reference using the base stack pointer. In the optimized version we have references to di erent registers such as %ebx and %edx.

How fast can we do this loop? Try using -O2 the second level of optimizations. What is happening now, is this really true? One should be careful when doing measurements, sometimes the quickest way to do something is not to do it at all.

2 The benchmark

So now for the actual benchmark. Were going to allocate a huge array that we will access in a linear fashion. The question is if it takes more time to access 4 pages compared to 16, 32 or even more pages.

2.1 a rst try

We rst de ne the page size and we assume that we have a page size of 64 Bytes. This is of course not the case since we know that the page size of our machine is probably 4 KiByte but we will use it as a rst step to see what it is that we are measuring.

#define PAGESIZE 64

Then we allocate a huge array and why not call it memory. Note that we cast the PAGESIZE to a long integer i order for the multiplication to work even if we start to use a total memory that is larger than 4GiByte (and we will in the end). We also run through all pages and write to them just to force the operating system to actually allocate the pages and set up the page table entries.

char memory = malloc (( long)PAGESIZE PAGES);

for ( int p = 0; p < PAGES; p++) {

long ref = (long)p PAGESIZE; / force the page to be allocated /

memory[ ref ] += 1;

}

We also add some nice printout so that we can look at our gures later and see what parameters we used. The initialization and printout is added before the code that starts the sequence of test.

printf (#TLB experiment
);

printf (# page size = %d bytes
, (PAGESIZE) ) ; printf (# max pages = %d
, (PAGES) ) ; printf (# total number or references = %d Mi
, (REFS/(10241024))); printf (#pagest proct sum
);

Then it is time to do the benchmark and we simply replace the dummy ( sum++) operation with a page reference that we use to increment the sum.

/ this w i l l replace our dummy operation / long ref = (long)p PAGESIZE;

sum += memory[ ref ] ; // sum++;

If you have everything in place you should see something like the printout below. You can compare these number to the dummy that we performed before (Ive used 10 million references).

#TLB experiment

# page size = 64 bytes

# max number of pages = 16

# total number or references = 10 Mi

# time for all references in sec (1000000)

#pages proc sum

1 0.040 10485760 2 0.060 10485760 3 0.043 10485759

4 0.036 10485760

:

:

There is a small di erence in accessing a memory references compared to just doing a dummy addition but the di erence is not very large. If you try using up to 64 pages you might notice something strange and predictable behaviour. We have sorted out why the initial tests show slightly higher values but there might also be other test that show higher values. The di erence likely has to do with the the size and structure of the rst level data caches.

The initial preparations of the benchmark are important since it tells us what order of magnitude the di erences have to be in order to detect them. If TLB misses only increase the execution time by 10 percent then we will have a hard time separating what is actually an e ect of TLB misses and what is an e ect data caches etc.

2.2 running some tests

Lets rst increase the page size to something more reasonable such as 4 KByte. This is probably the size that the machine that youre running on uses but if youre not running on a x86 architecture you should look it up.

#define PAGESIZE (41024)

First see if there is any di erence using 16, 32 or why not 64 pages. Hmm, switch back to a page size of 64 bytes and then try 64 pages again. How large is the memory that we use when we use 64 pages of size 64 byte each? How many real pages are used then?

What happens if we step up to 512 pages or why not 4096? Since were probably not that interested in what happens between 2378 pages and 2379 we can change the benchmark to try double as many pages in each iteration. for ( int pages = 4; pages <= PAGES; pages =2) {

Ouch, that hurts. There is de nitely a penalty that we take when using more pages.

2.3 turn it into a graph

So you now have some nice printouts of the time it takes to reference a page depending on the number of pages that we use. As you see there is de nitely a penalty as we use more pages. We can create a nice plot of the values using a tool called gnuplot.

First we save the output in a le called tlb.dat. This is easily done using the redirection operator of the shell.

$ gcc -o tlb -O tlb.c

$ ./tlb > tlb.dat

You can run gnuplot in interactive mode (and you should learn how to do so to quickly generate a graph) but its better to write a small script that generates the plot. Write the following in a le tlb.p. set terminal png set output tlb . png set t i t l e TLB benchmark , 4 KiByte pages , 10 Gi operations set key right center set xlabel number of pages set ylabel time in s

# use log scale if we use doubling of number of pages set logscale x 2 plot tlb . dat u 1:2 w linespoints t i t l e page refs

Now run gnuplot from the command line and let it execute the script.

> gnuplot tlb.p

You should now nd a le called tlb.png with a nice graph. Experiment with gnuplot and add more lines in the same graph. If you generate data for the dummy test, using pages of 64 bytes and 4K bytes you can print them in one graph using the following directive.

plot tlb4K . dat u 1:2 w linespoints t i t l e page size 4K bytes , tlb64 . dat u 1:2 w linespoints t i t l e page size 64 bytes , dummy. dat u 1:2 w linespoints t i t l e dummy

2.4 before you go to bed

One more experiment before youre done but make sure that your laptop is connected to power and dont sit up waiting. Were going to challenge the system and allocate as much memory as you have in your computer. We assume that youre running a 64-bit operating system and have something like 8 GiByte of memory. If you allocate 2 Mi pages, (2*1024*1024), youre asking for as much memory as you have. Since the operating system will need some memory it will have to swap pages to the hard-drive in order to meet your needs. This will take time . and Im not joking.

Check how much memory you have on your laptop and then set the number of pages so the whole memory would be needed. When you start the benchmark you can start it using the time command. This will after the execution report the wall time as well as the time used by the process. When the process is suspended waiting on a page to be brought in from disk the process time is not running. Start the benchmark and go to bed. In the morning you will be surprised how long time it took.