The WeakReference class, monitoring memory leak and garbage collection in a Java application

-
Image
 Below is a Stack implementation that uses an internal resizeable array structure.  public class MyStack< T > implements Stack< T > { private static final int CAPACITY = 100 ; private Object[] array ; private int pos = 0 ; public MyStack () { this . array = new Object[ CAPACITY ] ; } @Override public void push ( T item) { if ( pos >= array . length / 2 ) { Object[] newArray = new Object[ pos * 2 ] ; System. arraycopy ( array , 0 , newArray , 0 , array . length ) ; array = newArray ; } array [ pos ++] = item ; } @Override public T pop () { if (isEmpty()) { throw new RuntimeException( "empty stack" ) ; } @SuppressWarnings ( "unchecked" ) T item = ( T ) array [ pos - 1 ] ; pos -= 1 ; return item ; } @Override @SuppressWarnings ( "unchecked" ) public T peek...
Post a Comment

Returning to SimpleScalar and TLB

-
The topic of my master's degree thesis is Virtual Memory. More specifically, it is about a low-power Translation Lookaside Buffer design. You can look at my previous posts for an introduction.  

We have proven that more than 97% of virtual adress reading is done from the same page. So it is clear that the motivation of putting the TLB in low-power mode is a good idea if the cost of changing power mode is not too high.

The motivation part is ok. Now what I need is calculating the latency that is caused by putting the TLB in low-power mode. After that, I need to put there a mini-TLB, and simulate the situation. If the result is not better, then I'll say that mini-TLB is not required.

Note: Virtualbox->Settings->Display->Enable 3d acceleration

We check if the power mode is changed for every instruction:

 if (itlb) {

     if (CACHE_TAGSET(itlb, IACOMPRESS(fetch_regs_PC)) == itlb->last_tagset) {

     if (low_power == 0) {
                power_mode_changed = 1;
    } else {
                power_mode_changed = 0;
    }

     low_power = 1;
    ++low_power_count;

    } else {

             if (low_power == 1) {
                power_mode_changed = 1;
        } else {
                power_mode_changed = 0;
        }

             low_power = 0;
            ++normal_power_count;
    }

    tlb_lat = cache_access(itlb, Read, IACOMPRESS(fetch_regs_PC),
                     NULL, ISCOMPRESS(sizeof(md_inst_t)), sim_cycle,      NULL, NULL);


    if (power_mode_changed) {
           tlb_lat += 1;
   }

So if the power mode is changed, we add 1 clock cycle to the latency. Because according to Zhao et. al. we can use 1 cycle approximately. [1]


After running SPEC2000 programs with and without power-mode changes, I calculated the cost of changing power mode. Increase in total cycles depend on the latency caused by switching power mode.


GZIP

Latency: 1 cycle
normal:  580713144
power mode: 626491951
increment:

Latency: 10 cycle
normal: 580713144
power mode: 907701677
increment:

Latency: 25 cycle
normal: 580713144
power mode: 1388998717
increment:

Latency: 75 cycle
normal: 580713144
power mode: 2994729275
increment:

Latency: 100 cycles
normal: 580713144
power mode: 3797723125
increment: %18



crafty: latency=100   cycle: 5507522361
latency=75  cycle: 4325626411
latency=25  cycle: 1963406203
latency=10  cycle: 1256704284
latency=1   cycle: 864566362
latency=0   cycle: 800535108

parser: latency=0  cycle: 651830989
    latency=1  cycle: 682037147
    latency=10 cycle: 964905768
    latency=25 cycle: 1483275240
    latency=75 cycle: 3228328486
    latency=100 cycle: 4101332934


MESA

Latency: 1 cycle
normal: 706604
power mode: 710210
increment: % 0.5

Latency: 10 cycle
normal: 706604
power mode: 873556
increment: %23

Latency: 25 cycle
normal: 706604
power mode: 1153888
increment: %63

Latency: 75 cycle
normal: 706604
power mode: 2092983
increment: %196

Latency: 100 cycles
normal: 706604
power mode: 2562917
increment: %262

Latency: 8000 cycles
total cycle: 706570
power mode: 151051283
increment:  %212

MCF

Latency: 1 cycle
normal: 16270
power mode: 17236
increment: %5

Latency: 10 cycles
normal: 16270
power mode: 21796
increment: %33

Latency: 25 cycles
normal: 16270
power mode: 29826
increment: %83

Latency: 75 cycles
normal: 16270
power mode: 58668
increment: %260

Latency: 100 cycles
normal: 16270
power mode: 73243
increment: %350

BZIP2

Latency: 1 cycle
normal: 562851
power mode: 564804
increment:%0.34

Latency: 10 cycles
normal: 562851
power mode: 573788
increment: %1.9

Latency: 25 cycles
normal: 562851
power mode: 590123
increment: %4.8

Latency: 75 cycles
normal: 562851
power mode: 647554
increment: %15

Latency: 100 cycles
normal: 562851
power mode: 676534
increment: %20



PARSER

Latency: 1 cycles
normal: 15593
power mode: 16596
increment: %6

Latency: 10 cycles
normal: 15593
power mode: 20507
increment: %31

Latency: 25 cycles
normal: 15593
power mode: 27416
increment: %75

Latency: 75 cycles
normal: 15593
power mode: 52106
increment: %234


Latency: 100 cycles
normal: 15593
power mode: 64581
increment: %314


VPR

Latency: 1 cycle
normal: 19343
power mode: 20829
increment: %7

Latency: 10 cycle
normal: 19343
power mode: 27955
increment: %44

Latency: 25 cycle
normal: 19343
power mode: 40560
increment: %109

Latency: 75 cycle
normal: 19343
power mode: 83936
increment: %333

Latency: 100 cycles
normal: 19343
power mode: 105786
increment: %446

EQUAKE

Latency: 1 cycle
normal: 22668
power mode: 24636
increment: %8

Latency: 10 cycles
normal: 22668
power mode: 34961
increment: %54

Latency: 25 cycles
normal: 22668
power mode: 52939
increment: %133

Latency: 75 cycles
normal: 22668
power mode:  114399
increment: %404

Latency: 100 cycles
normal: 22668
power mode: 145374
increment: %541



[1] Zhao, L. E. I., et al. "A leakage efficient instruction tlb design for embedded processors." IEICE TRANSACTIONS on Information and Systems 94.8 (2011): 1565-1574.

Comments

Post a Comment

Popular posts from this blog

Trie Data Structure and Finding Patterns in a Collection of Words

-
Image
 I faced a very hard problem recently about finding substrings based on a collection of patterns.  Example For  words = ["Apple", "Melon", "Orange", "Watermelon"]  and  parts = ["a", "mel", "lon", "el", "An"] , the output should be  findSubstrings(words, parts) = ["Apple", "Me[lon]", "Or[a]nge", "Water[mel]on"] . While  "Watermelon"  contains three substrings from the  parts  array,  "a" ,  "mel" , and  "lon" ,  "mel"  is the longest substring that appears first in the string. Note: Number of words and number of parts can be huge! A brute force method or careless usage of Java String methods would make it impossible to handle a large number of words and parts (patterns). Solution Approach Whenever we want to search patterns inside a string, we should think about the Trie data structure:  https://en.wikipedia....
Read more

My Crappy Looking Solution to "Binary Tree Common Ancestor" Problem

-
I was able to solve this problem and make all the tests pass: https://www.hackerrank.com/challenges/binary-search-tree-lowest-common-ancestor/problem But I'm not happy with the code, so later I need to refactor the code. Here it is: static Node lca(Node root, int v1, int v2) { ArrayList result = new ArrayList<>(); Node result2 = traverse (root, v1, v2, result); if (result.isEmpty()) return result2; return result.get( 0 ); } private static Node traverse(Node root, int v1, int v2, ArrayList result) { if (!result.isEmpty()) { return result.get( 0 ); } if (root. left == null && root. right == null ) { return root; } if (root. data == v1 || root. data == v2) { return root; } Node node1 = null , node2 = null ; if (root. left != null ) node1 = traverse (root. left , v1, v2, result); if (root. right != null ) node2 = traverse (root. right , v1, v...
Read more

A Graph Application in Java: Using WordNet to Find Outcast Words

-
Image
I'll explain our intent with an example: Let's say we have the following words: horse, zebra, cat, bear, table Which one is the outcast? The table.. But how can a computer system find this? We are talking about the whole English language and any word can be in this list. The number of words in the list can be huge, so the solution should use appropriate data structures and algorithms.  We'll use WordNet which is a graph of English words. It contains the semantic relationships between words and our concern is the "is a" relationship. For our example, we know that horse and cat are both animals. "Table" is not an animal and it is the outcast in this situation. But we even have the ability to tell that horse and zebra is very close. Below is a unit test that shows a high level view of our application in Java: public class OutcastTest {     private WordNet wordNet;     private Outcast outcast;     @Before ...
Read more