Program Specialization of
Systems Programs
Charles Consel
Phoenix Research Group
(formerly known as the Compose Group)
INRIA -LaBRI
October 2005
1
INRIA - LaBRI
Phoenix Group
Oct-05
Phoenix Group: Research Topics

Programming language technology
–
–
–

Application areas
–
–
–

Operating systems
Networking
Telecommunications
Prototypes
–
–
2
Program analysis and transformation
Program specialization
Domain-specific languages
Tempo
Devil, Plan-P, Spidle, SPL
Phoenix Group
Oct-05
INRIA - LaBRI
Recent Contributors

Anne-Françoise Le Meur, University of Lille
Sapan Bhatia, PhD student in Phoenix
Julia Lawall, University of Copenhagen
3
Phoenix Group


Oct-05
INRIA - LaBRI
Adaptable Programs:
Why, Where, How
Introduction
4
INRIA - LaBRI
Phoenix Group
Oct-05
Adaptable Programs: Why?
Program1
Problem 2
Program2
Problem 3
Program3
Adaptable
Program
…
Problem 1
Problem
Tool
Adapted
Program
Problem family
5
Program family
Phoenix Group
Oct-05
INRIA - LaBRI
Adaptable Programs: Where?

Areas:
–
–

Features:
–
–
–
–
6
Operating system components
Networking layers
Adaptability to a variety of platforms
Adaptability to a family of services
Adaptability to a variety of states
Long-running programs
Phoenix Group
Oct-05
INRIA - LaBRI
Adaptable Programs: How?
Program structuring approach: generic programs
Generic
Program
Customization
information
Specializer
Customized
Program
7
Phoenix Group
Oct-05
INRIA - LaBRI
Specialization of System/Networking Code:
Challenges

System code
–
–
–

Program specialization
–
–
–
–
8
Legacy usually written in C
Generic but optimized code
Poor software architecture
Specialization of a real-size language (i.e., C)
Specialization of a low-level language
Specialization at run time (not only compile time)
Targeting real-size cases
Phoenix Group
Oct-05
INRIA - LaBRI
Specializing System/Networking Code

Synthesis Kernel - Columbia
–

Synthetix Project – OGI
–

Ad Hoc manual run-time code generation for system calls
[SOSP’89]
Methodology for manual specialization [SOSP’95]
Tempo - Compose/OGI

Automatic specialization [TOCS’01]


9
Sun Remote Procedure Call [ICDCS’98]
Incremental Checkpointing [DSN’00]
Phoenix Group
Oct-05
INRIA - LaBRI
Tempo in One Slide!


Specialization declarations
Program analyses

Context/flow/return sensitive
binding-time analysis
Evaluation-time analysis
Action analysis

–
–
–



–
Program transformations
–
–
–
Compile time
Run time
Data specialization
Program specializer:
Tempo
Languages: C, C++, Java
Software architectures
Components:
–
–
–

IPC
RPC
Signals
TCP/IP
Gains: time (and/or) space
[…SOSP’95, POPL’96, PLDI’99, ECOOP’99, ASE’00, CD’02, EMSOFT’04, LCN’04…]
[…HOSC’99, HOSC’00, TOCS’01, TOPLAS’03, HOSC’04, SCP’04…]
10
Phoenix Group
Oct-05
INRIA - LaBRI
Anatomy of Tempo
Analysis ctx
C program
Preprocessing
Specialization ctx
Compile-time
Specializer
Postprocessing
Run-time specializer
Generator
Run-time
Specializer
Specialization ctx
Specialized source
Specialized binary
11
Phoenix Group
Oct-05
INRIA - LaBRI
Anatomy of Tempo
Analysis ctx
C program
Preprocessing
Specialization ctx
Compile-time
Specializer
Postprocessing
Run-time specializer
Generator
Run-time
Specializer
Specialization ctx
Specialized source
Specialized binary
12
Phoenix Group
Oct-05
INRIA - LaBRI
Integrating Program Specialization in
The Software Development Process
Part I
13
INRIA - LaBRI
Phoenix Group
Oct-05
Context

Customizable component  Family of problems
Customizable
for a family of
problems
Customization
information
Component
?
Description of a
given problem
(a usage context)
Component customization
Customized
component
14
Phoenix Group
Oct-05
INRIA - LaBRI
Software Component Customization

Functional customization
–
restriction of the behavior


ex: JavaBeans
Code customization
–
restriction of the behavior
–
reduction of the genericity
smaller & faster code
15
Phoenix Group
Oct-05
INRIA - LaBRI
Code customization

Scope :
–

often program fragments
How :
–
ifdef compiler directives

–
C++ Templates


cumbersome, error-prone, difficult to use, no debugger
Limitation :
–
16
code hard to read, hard to maintain, complex
the will of the programmer to encode the transformations
Phoenix Group
Oct-05
INRIA - LaBRI
Example: power
int power(int base, int expon) {
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ; }
return accum; }
17
Phoenix Group
Oct-05
INRIA - LaBRI
Example: power
int power(int base, int expon) {
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ; }
return accum; }
Functional customization
18
Phoenix Group
Oct-05
INRIA - LaBRI
Example: power
int power(int base, int expon) {
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ; }
return accum; }
int power(int base, 3) {
int accum = 1;
while ( expon > 0) {
accum *= base;
expon - - ; }
return accum; }
19
Functional customization
Phoenix Group
Oct-05
INRIA - LaBRI
Example: power
int power(int base, int expon) {
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ; }
return accum; }
Code customization
20
Phoenix Group
Oct-05
INRIA - LaBRI
Example: power
int power(int base, int expon) {
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ; }
return accum; }
expon = 3
Code customization
int power(int base){
return base*base*base;
}
21
Phoenix Group
Oct-05
INRIA - LaBRI
Example: power
int power(int base, int expon) {
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ; }
return accum; }
expon = 3
int power(int base){
return base*base*base;
}
22
Phoenix Group
template<int expon>
inline int power(const int& base)
{ return power<expon-1>(base) * base;}
template<>
inline int power<1>(const int& base)
{ return base;}
template<>
inline int power<0>(const int& base)
{ return 1;}
Oct-05
INRIA - LaBRI
Example: power
int power(int base, int expon) {
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ; }
return accum; }
expon = 3
int power(int base){
return base*base*base;
}
23
Phoenix Group
template<int expon>
inline int power(const int& base)
{ return power<expon-1>(base) * base;}
template<>
inline int power<1>(const int& base)
{ return base;}
template<>
inline int power<0>(const int& base)
{ return 1;}
Poor readability
Oct-05
INRIA - LaBRI
Example: power
int power(int base, int expon) {
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ; }
return accum; }
expon = 3
int power(int base){
return base*base*base;
}
24
Phoenix Group
template<int expon>
inline int power(const int& base)
{ return power<expon-1>(base) * base;}
template<>
inline int power<1>(const int& base)
{ return base;}
template<>
inline int power<0>(const int& base)
{ return 1;}
Semantic transformation
Oct-05
INRIA - LaBRI
Example: power
int power(int base, int expon) {
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ; }
return accum; }
expon = 3
int power(int base){
return base*base*base;
}
template<int expon>
inline int power(const int& base)
{ return power<expon-1>(base) * base;}
template<>
inline int power<1>(const int& base)
{ return base;}
template<>
inline int power<0>(const int& base)
{ return 1;}
power<3>
25
Phoenix Group
Oct-05
INRIA - LaBRI
Example: power
int power(int base, int expon) {
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ; }
return accum; }
expon = 3
int power(int base){
return base*base*base;
}
template<int expon>
inline int power(const int& base)
{ return power<expon-1>(base) * base;}
template<>
inline int power<1>(const int& base)
{ return base;}
template<>
inline int power<0>(const int& base)
{ return 1;}
power<2>(base) * base
26
Phoenix Group
Oct-05
INRIA - LaBRI
Example: power
int power(int base, int expon) {
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ; }
return accum; }
expon = 3
int power(int base){
return base*base*base;
}
template<int expon>
inline int power(const int& base)
{ return power<expon-1>(base) * base;}
template<>
inline int power<1>(const int& base)
{ return base;}
template<>
inline int power<0>(const int& base)
{ return 1;}
power<1>(base) * base * base
27
Phoenix Group
Oct-05
INRIA - LaBRI
Example: power
int power(int base, int expon) {
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ; }
return accum; }
expon = 3
int power(int base){
return base*base*base;
}
template<int expon>
inline int power(const int& base)
{ return power<expon-1>(base) * base;}
template<>
inline int power<1>(const int& base)
{ return base;}
template<>
inline int power<0>(const int& base)
{ return 1;}
base * base * base
28
Phoenix Group
Oct-05
INRIA - LaBRI
Example: power
int power(int base, int expon) {
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ; }
return accum; }
expon = 3
int power(int base){
return base*base*base;
}
template<int expon>
inline int power(const int& base)
{ return power<expon-1>(base) * base;}
template<>
inline int power<1>(const int& base)
{ return base;}
template<>
inline int power<0>(const int& base)
{ return 1;}
Do we really get what we wanted ??
base * base * base
29
Phoenix Group
Oct-05
INRIA - LaBRI
To Summarize
Customization
information
Component
?
Customized
component
30
Phoenix Group
Oct-05
INRIA - LaBRI
To Summarize
Source code
with
hardcoded
transformations :
what to customize
& how
Customization
information
Component
?
Customized
component
31
Phoenix Group
Oct-05
INRIA - LaBRI
To Summarize
Source code
with
hardcoded
transformations :
what to customize
& how
Customization
information
Component
Values
?
Customized
component
32
Phoenix Group
Oct-05
INRIA - LaBRI
To Summarize
Source code
with
hardcoded
transformations :
what to customize
& how
Customization
information
Component
Values
Compiler
Customized
component
33
Phoenix Group
Oct-05
INRIA - LaBRI
To Summarize
Source code
with
hardcoded
transformations :
what to customize
& how
Customization
information
Component
Compiler
Customized
component
34
Values
Phoenix Group
Oct-05
No real guaranty
INRIA - LaBRI
What We Would Like
Source code
with
hardcoded
transformations :
what to customize
& how
Customization
information
Component
Compiler
Customized
component
35
Values
Phoenix Group
Oct-05
No real guaranty
INRIA - LaBRI
What We Would Like
Source
code
What to
customize
Customization
information
Values
Compiler
Customized
component
36
Phoenix Group
Oct-05
INRIA - LaBRI
What We Would Like
Source
code
37
What to
customize
Customization
information
Automatic generation
of hardcoded
transformations
Compiler
Code
with annotated
transformations
Customized
component
Phoenix Group
Oct-05
Values
INRIA - LaBRI
What We Would Like
Source
code
38
What to
customize
Customization
information
Automatic generation
of hardcoded
transformations
Compiler
Code
with annotated
transformations
Customized
component
Phoenix Group
Oct-05
Values
INRIA - LaBRI
What We Would Like
Source
code
39
What to
customize
Customization
information
Automatic generation
of hardcoded
transformations
Compiler
Code
with annotated
transformations
Customized
component
Phoenix Group
Oct-05
Values
Guaranteed
INRIA - LaBRI
What We Would Like
Source
code
40
Values
What to
customize
Customization
information
Automatic generation
of hardcoded
transformations
Compiler
Code
with annotated
transformations
Customized
component
Phoenix Group
Oct-05
Guaranteed
INRIA - LaBRI
What We Would Like
Source
code
Values
What to
customize
Customization
information
Automatic generation
of hardcoded
transformations
Compiler
Code
with annotated
transformations
Customized
component
Guaranteed
Developer
41
Phoenix Group
Oct-05
INRIA - LaBRI
What We Would Like
Source
code
Values
What to
customize
Customization
information
Automatic generation
of hardcoded
transformations
Compiler
Code
with annotated
transformations
Customized
component
Developer
42
Guaranteed
User
Phoenix Group
Oct-05
INRIA - LaBRI
Our Approach
Component developer
Source
code
Component user
Customization
scenarios
Customizable
component
Customizable
component
43
Customization
values
Customized
component
Phoenix Group
Oct-05
INRIA - LaBRI
Component Developer
Source
code
Customization
scenarios
Customizable
component
44
Phoenix Group
Oct-05
INRIA - LaBRI
Component Developer
Source
code
Step 1: Developing
customizable components
Customization
scenarios
Customizable
component
45
Phoenix Group
Oct-05
INRIA - LaBRI
Component Developer
Source
code
Step 1: Developing
customizable components
Customization
scenarios
Declaration language: WHAT
are the customization parameters
• global variables
• function parameters
• data structure fields
Customizable
component
46
Phoenix Group
Oct-05
INRIA - LaBRI
An example of component :
Forward Error Correction Encoders
• Prevent losses and errors
• Transmit redundant information
family of problems
47
Phoenix Group
Oct-05
INRIA - LaBRI
Encoder Features
48
Phoenix Group
Oct-05
INRIA - LaBRI
FEC Component Hierarchy
encode
parity bit
_
lbc sys
_
code conv
_
lbc
crc
libstring
mult mat
_
49
Phoenix Group
Oct-05
INRIA - LaBRI
Remember power
source code
int power(int base, int expon)
{
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ;
}
return accum;
}
50
Phoenix Group
Oct-05
INRIA - LaBRI
Remember power
source code
int power(int base, int expon)
{
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ;
}
return accum;
}
51
Phoenix Group
Oct-05
INRIA - LaBRI
Remember power
source code
declaration module
int power(int base, int expon)
{
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ;
}
return accum;
}
52
Module power {
Defines {
From power.c {
Btp :: intern power(D(int) b, S(int) e);
};}
Exports { Btp; }}
Phoenix Group
customization parameters
other parameters
Oct-05
INRIA - LaBRI
Remember power
source code
declaration module
int power(int base, int expon)
{
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ;
}
return accum;
}
53
Module power {
Defines {
From power.c {
Btp :: intern power(D(int) b, S(int) e);
};}
Exports { Btp; }}
Phoenix Group
customization parameters
other parameters
Oct-05
INRIA - LaBRI
Remember power
source code
declaration module
int power(int base, int expon)
{
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ;
}
return accum;
}
54
Module power {
Defines {
From power.c {
Btp :: intern power(D(int) b, S(int) e);
};}
Exports { Btp; }}
Phoenix Group
customization parameters
other parameters
Oct-05
INRIA - LaBRI
Remember power
source code
declaration module
int power(int base, int expon)
{
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ;
}
return accum;
}
55
Module power {
Defines {
From power.c {
Btp :: intern power(D(int) b, S(int) e);
};}
Exports { Btp; }}
Phoenix Group
customization parameters
other parameters
Oct-05
INRIA - LaBRI
Remember power
source code
declaration module
int power(int base, int expon)
{
int accum = 1;
while (expon > 0) {
accum *= base;
expon - - ;
}
return accum;
}
56
Module power {
Defines {
From power.c {
Btp :: intern power(D(int) b, S(int) e);
};}
Exports { Btp; }}
Phoenix Group
customization parameters
other parameters
Oct-05
INRIA - LaBRI
Remember power
Define: compute
needs to perform power calculation
Define: power
57
Phoenix Group
Oct-05
Component:
compute
Component:
power
INRIA - LaBRI
Remember power
Module compute
{ Imports { From power.mdl {Btp; }}
Defines {
From compute.c {
Btcomp :: intern compute(D(int) , S(int) , S(int) )
needs { Btp;} };}
Exports { Btcomp; } }
Module power
{ Defines {
From power.c {
Btp :: intern power(D(int) , S(int) );
};}
Exports { Btp; } }
58
Phoenix Group
Oct-05
Module:
compute
Module:
power
INRIA - LaBRI
Remember power
Module compute
{ Imports { From power.mdl {Btp; }}
Defines {
From compute.c {
Btcomp :: intern compute(D(int) , S(int) , S(int) )
needs { Btp;} };}
Exports { Btcomp; } }
Module:
compute
Customization contract
Module power
{ Defines {
From power.c {
Btp :: intern power(D(int) , S(int) );
};}
Exports { Btp; } }
59
Phoenix Group
Oct-05
Module:
power
INRIA - LaBRI
Hierarchies of Sources & Modules
encode
parity bit
_
lbc sys
_
code conv
_
lbc
crc
libstring
mult mat
_
60
Phoenix Group
Oct-05
INRIA - LaBRI
Developing Customizable Components
61
Phoenix Group
Oct-05
INRIA - LaBRI
Developing Customizable Components
62
Phoenix Group
Oct-05
INRIA - LaBRI
Developing Customizable Components
63
Phoenix Group
Oct-05
INRIA - LaBRI
Developing Customizable Components
64
Phoenix Group
Oct-05
INRIA - LaBRI
Developing Customizable Components
65
Phoenix Group
Oct-05
INRIA - LaBRI
Developing Customizable Components
66
Phoenix Group
Oct-05
INRIA - LaBRI
Developing Customizable Components
67
Phoenix Group
Oct-05
INRIA - LaBRI
Developing Customizable Components
68
Phoenix Group
Oct-05
INRIA - LaBRI
Component Developer
Source
code
Step 1: Developing
customizable components
Customization
scenarios
Customizable
component
69
Phoenix Group
Oct-05
INRIA - LaBRI
Component Developer
Source
code
Step 1: Developing
customizable components
Customization
scenarios
Step 2: Making
customizable components
Customizable
component
70
Phoenix Group
Oct-05
INRIA - LaBRI
Making Customizable Components
71
Phoenix Group
Oct-05
INRIA - LaBRI
Component Developer
Source
code
Step 1: Developing
customizable components
Customization
scenarios
Step 2: Making
customizable components
- Determine HOW
- Verify that declared
customization is possible
Customizable
component
72
Phoenix Group
Oct-05
INRIA - LaBRI
Component Developer
Source
code
Step 1: Developing
customizable components
Customization
scenarios
Step 2: Making
customizable components
Customizable
component
73
Guaranteed
Phoenix Group
Oct-05
INRIA - LaBRI
Our Approach
Component developer
Source
code
Component user
Customization
scenarios
Customizable
component
Customizable
component
74
Customization
values
Customized
component
Phoenix Group
Oct-05
INRIA - LaBRI
Component User
Customizable
component
Customization
values
Customized
component
75
Phoenix Group
Oct-05
INRIA - LaBRI
Component User
Customizable
component
Customization
values
Step 3: Making
customized components
Customized
component
76
Phoenix Group
Oct-05
INRIA - LaBRI
Generating Customized Components
77
Phoenix Group
Oct-05
INRIA - LaBRI
Generating Customized Components
78
Phoenix Group
Oct-05
INRIA - LaBRI
Generating Customized Components
79
Phoenix Group
Oct-05
INRIA - LaBRI
Component User
Customizable
component
Customization
values
Step 3: Making
customized components
Customized
component
80
Phoenix Group
Oct-05
INRIA - LaBRI
Component User
Customizable
component
Customization
values
Step 3: Making
customized components
Ready to be integrated
81
Phoenix Group
Oct-05
Customized
component
INRIA - LaBRI
#include
#include
#include
#include
#include
#include
#include
"encode.h"
"parity_bit.h"
"lin_b_code.h"
"lin_b_code_sys.h"
"code_conv.h"
"encode_crc.h"
"libstring.h"
struct data
{
int longueur;
int *v;
};
void
encode(struct data *donnees, struct code *codage, struct data *result){
struct code
if (donnees->longueur != codage->k)
{
{ perror("donnees->longueur != codage->k");
int k;
exit(1);}
int n;
else if (strcmp(codage->type, "parity_bit") == 0)
int **matrix;
if (result->longueur != codage->k + 1)
char *type;
{ perror("result->longueur != codage->k + 1");
char *opt;
exit(1);}
};
else parity_bit(codage->k, donnees->v, result->v);
else if (strcmp(codage->type, "lin_b_code") == 0)
extern void encode(struct data *donnees, struct code *codage, struct data *result);
if ((*result).longueur != (*codage).n)
{ perror ("result->longueur != codage->n");
exit (1);}
else if (strcmp(codage->opt, "syst") == 0)
lin_b_code_sys(donnees->v, codage->k, codage->n, codage->matrix, result->v);
else
lin_b_code(donnees->v, codage->k, codage->n, codage->matrix, result->v);
else if (strcmp(codage->type, "code_conv") == 0)
if ((*result).longueur != (2 * (*codage).k))
{ perror ("result->longueur != (2 * codage->k)");
exit (1);}
else code_conv(donnees->v, codage->k, result->v);
else if (strcmp(codage->type, "crc") == 0)
if ((*result).longueur != (*codage).n )
{ perror ("result->longueur != codage->n");
exit (1); }
else encode_crc(donnees->v, codage->k, codage->n, codage->matrix[0], result->v);
else
{ perror("unknown code type");
extern int strcmp(char *s1, char *s2);
exit(1); }
}
extern void encode_crc(int *donnees, int k, int n, int *poly, int *result);
extern void parity_bit(int l, int *vector, int *result);
#include "parity_bit.h"
void
parity_bit(int l, int *vector, int *result)
{
int res;
int i;
res = 0;
for(i = 0; i < l; i++)
{
res ^= vector[i];
result[i] = vector[i];
}
result[l] = res;
extern void code_conv(int *donnees, int l, int *result);
extern void lin_b_code_sys(int *vector, int k, int n, int **matrix, int *result);
#include "code_conv.h"
void
code_conv(int *donnees, int k, int *result)
{
int temp;
int i;
#include "lin_b_code_sys.h"
#include "mult_mat.h"
void
lin_b_code_sys(int *vector, int k, int n, int **matrix, int *result)
{
int sub_matrix_size;
int i;
for (i = 0; i < k - 2; i++)
{
temp = donnees[i] ^ donnees[i + 2];
result[2 * i + 1] = temp;
temp ^= donnees[i + 1];
result[2 * i] = temp;
}
temp = donnees[k - 2] ^ donnees[k - 1];
result[2 * k - 4] = temp;
result[2 * k - 3] = donnees[k - 2];
result[2 * k - 2] = donnees[k - 1];
result[2 * k - 1] = donnees[k - 1];
sub_matrix_size = n - k;
mult_mat(vector, k, sub_matrix_size, matrix, result, k);
for(i = 0; i < k; i++)
result[i] = vector[i];
}
}
}
#include "encode_crc.h"
void
encode_crc(int *donnees, int k, int n, int *poly, int *result)
{
int l_poly;
int i, j;
l_poly = n - k + 1;
for (i = 0; i < k; i++)
result[i] = donnees[i];
for (i = 0; i < k; i++)
if (result[i] == 1)
for (j = 0; j < l_poly; j++)
result[i + j] ^= poly[j];
for (i = 0; i < k; i++)
result[i] = donnees[i];
}
extern void lin_b_code(int *vector, int k, int n, int **matrix, int *result);
#include "lin_b_code.h"
#include "mult_mat.h"
void
lin_b_code(int *vector, int k, int n, int **matrix, int *result)
{
mult_mat(vector, k, n, matrix, result, 0);
}
extern void mult_mat(int *vector, int k, int n, int **matrix, int *result, int indice);
#include "mult_mat.h"
void
mult_mat(int *vector, int k, int n, int **matrix, int *result, int indice)
{
int temp;
int i, j;
for(i = 0; i < n; i++)
{
temp = 0;
for(j = 0; j < k; j++)
temp ^= vector[j] & matrix[j][i];
result[i + indice] = temp;
82
}
}
Phoenix Group
Oct-05
INRIA - LaBRI
#include
#include
#include
#include
#include
#include
#include
length in
length out
type
generator
opt
"encode.h"
"parity_bit.h"
"lin_b_code.h"
"lin_b_code_sys.h"
"code_conv.h"
"encode_crc.h"
"libstring.h"
struct data
{
int longueur;
int *v;
};
void
encode(struct data *donnees, struct code *codage, struct data *result){
struct code
if (donnees->longueur != codage->k)
{
{ perror("donnees->longueur != codage->k");
int k;
exit(1);}
int n;
else if (strcmp(codage->type, "parity_bit") == 0)
int **matrix;
if (result->longueur != codage->k + 1)
char *type;
{ perror("result->longueur != codage->k + 1");
char *opt;
exit(1);}
};
else parity_bit(codage->k, donnees->v, result->v);
else if (strcmp(codage->type, "lin_b_code") == 0)
extern void encode(struct data *donnees, struct code *codage, struct data *result);
if ((*result).longueur != (*codage).n)
{ perror ("result->longueur != codage->n");
exit (1);}
else if (strcmp(codage->opt, "syst") == 0)
lin_b_code_sys(donnees->v, codage->k, codage->n, codage->matrix, result->v);
else
lin_b_code(donnees->v, codage->k, codage->n, codage->matrix, result->v);
else if (strcmp(codage->type, "code_conv") == 0)
if ((*result).longueur != (2 * (*codage).k))
{ perror ("result->longueur != (2 * codage->k)");
exit (1);}
else code_conv(donnees->v, codage->k, result->v);
else if (strcmp(codage->type, "crc") == 0)
if ((*result).longueur != (*codage).n )
{ perror ("result->longueur != codage->n");
exit (1); }
else encode_crc(donnees->v, codage->k, codage->n, codage->matrix[0], result->v);
else
{ perror("unknown code type");
extern int strcmp(char *s1, char *s2);
exit(1); }
}
extern void encode_crc(int *donnees, int k, int n, int *poly, int *result);
extern void parity_bit(int l, int *vector, int *result);
#include "parity_bit.h"
void
parity_bit(int l, int *vector, int *result)
{
int res;
int i;
res = 0;
for(i = 0; i < l; i++)
{
res ^= vector[i];
result[i] = vector[i];
}
result[l] = res;
extern void code_conv(int *donnees, int l, int *result);
extern void lin_b_code_sys(int *vector, int k, int n, int **matrix, int *result);
#include "code_conv.h"
void
code_conv(int *donnees, int k, int *result)
{
int temp;
int i;
#include "lin_b_code_sys.h"
#include "mult_mat.h"
void
lin_b_code_sys(int *vector, int k, int n, int **matrix, int *result)
{
int sub_matrix_size;
int i;
for (i = 0; i < k - 2; i++)
{
temp = donnees[i] ^ donnees[i + 2];
result[2 * i + 1] = temp;
temp ^= donnees[i + 1];
result[2 * i] = temp;
}
temp = donnees[k - 2] ^ donnees[k - 1];
result[2 * k - 4] = temp;
result[2 * k - 3] = donnees[k - 2];
result[2 * k - 2] = donnees[k - 1];
result[2 * k - 1] = donnees[k - 1];
sub_matrix_size = n - k;
mult_mat(vector, k, sub_matrix_size, matrix, result, k);
for(i = 0; i < k; i++)
result[i] = vector[i];
}
}
}
#include "encode_crc.h"
void
encode_crc(int *donnees, int k, int n, int *poly, int *result)
{
int l_poly;
int i, j;
l_poly = n - k + 1;
for (i = 0; i < k; i++)
result[i] = donnees[i];
for (i = 0; i < k; i++)
if (result[i] == 1)
for (j = 0; j < l_poly; j++)
result[i + j] ^= poly[j];
for (i = 0; i < k; i++)
result[i] = donnees[i];
}
extern void lin_b_code(int *vector, int k, int n, int **matrix, int *result);
#include "lin_b_code.h"
#include "mult_mat.h"
void
lin_b_code(int *vector, int k, int n, int **matrix, int *result)
{
mult_mat(vector, k, n, matrix, result, 0);
}
extern void mult_mat(int *vector, int k, int n, int **matrix, int *result, int indice);
#include "mult_mat.h"
void
mult_mat(int *vector, int k, int n, int **matrix, int *result, int indice)
{
int temp;
int i, j;
for(i = 0; i < n; i++)
{
temp = 0;
for(j = 0; j < k; j++)
temp ^= vector[j] & matrix[j][i];
result[i + indice] = temp;
83
}
}
Phoenix Group
Oct-05
INRIA - LaBRI
#include
#include
#include
#include
#include
#include
#include
length in
length out
type
generator
opt
"encode.h"
"parity_bit.h"
"lin_b_code.h"
"lin_b_code_sys.h"
"code_conv.h"
"encode_crc.h"
"libstring.h"
struct data
{
int longueur;
int *v;
};
void
encode(struct data *donnees, struct code *codage, struct data *result){
struct code
if (donnees->longueur != codage->k)
{
{ perror("donnees->longueur != codage->k");
int k;
exit(1);}
int n;
else if (strcmp(codage->type, "parity_bit") == 0)
int **matrix;
if (result->longueur != codage->k + 1)
char *type;
{ perror("result->longueur != codage->k + 1");
char *opt;
exit(1);}
};
else parity_bit(codage->k, donnees->v, result->v);
else if (strcmp(codage->type, "lin_b_code") == 0)
extern void encode(struct data *donnees, struct code *codage, struct data *result);
if ((*result).longueur != (*codage).n)
{ perror ("result->longueur != codage->n");
exit (1);}
else if (strcmp(codage->opt, "syst") == 0)
lin_b_code_sys(donnees->v, codage->k, codage->n, codage->matrix, result->v);
else
lin_b_code(donnees->v, codage->k, codage->n, codage->matrix, result->v);
else if (strcmp(codage->type, "code_conv") == 0)
if ((*result).longueur != (2 * (*codage).k))
{ perror ("result->longueur != (2 * codage->k)");
exit (1);}
else code_conv(donnees->v, codage->k, result->v);
else if (strcmp(codage->type, "crc") == 0)
if ((*result).longueur != (*codage).n )
{ perror ("result->longueur != codage->n");
exit (1); }
else encode_crc(donnees->v, codage->k, codage->n, codage->matrix[0], result->v);
else
{ perror("unknown code type");
extern int strcmp(char *s1, char *s2);
exit(1); }
}
extern void encode_crc(int *donnees, int k, int n, int *poly, int *result);
extern void parity_bit(int l, int *vector, int *result);
#include "parity_bit.h"
void
parity_bit(int l, int *vector, int *result)
{
int res;
int i;
res = 0;
for(i = 0; i < l; i++)
{
res ^= vector[i];
result[i] = vector[i];
}
result[l] = res;
extern void code_conv(int *donnees, int l, int *result);
extern void lin_b_code_sys(int *vector, int k, int n, int **matrix, int *result);
#include "code_conv.h"
void
code_conv(int *donnees, int k, int *result)
{
int temp;
int i;
#include "lin_b_code_sys.h"
#include "mult_mat.h"
void
lin_b_code_sys(int *vector, int k, int n, int **matrix, int *result)
{
int sub_matrix_size;
int i;
for (i = 0; i < k - 2; i++)
{
temp = donnees[i] ^ donnees[i + 2];
result[2 * i + 1] = temp;
temp ^= donnees[i + 1];
result[2 * i] = temp;
}
temp = donnees[k - 2] ^ donnees[k - 1];
result[2 * k - 4] = temp;
result[2 * k - 3] = donnees[k - 2];
result[2 * k - 2] = donnees[k - 1];
result[2 * k - 1] = donnees[k - 1];
sub_matrix_size = n - k;
mult_mat(vector, k, sub_matrix_size, matrix, result, k);
for(i = 0; i < k; i++)
result[i] = vector[i];
}
}
}
#include "encode_crc.h"
void
encode_crc(int *donnees, int k, int n, int *poly, int *result)
{
int l_poly;
int i, j;
l_poly = n - k + 1;
for (i = 0; i < k; i++)
result[i] = donnees[i];
for (i = 0; i < k; i++)
if (result[i] == 1)
for (j = 0; j < l_poly; j++)
result[i + j] ^= poly[j];
for (i = 0; i < k; i++)
result[i] = donnees[i];
}
extern void lin_b_code(int *vector, int k, int n, int **matrix, int *result);
#include "lin_b_code.h"
#include "mult_mat.h"
void
lin_b_code(int *vector, int k, int n, int **matrix, int *result)
{
mult_mat(vector, k, n, matrix, result, 0);
}
extern void mult_mat(int *vector, int k, int n, int **matrix, int *result, int indice);
#include "mult_mat.h"
void
mult_mat(int *vector, int k, int n, int **matrix, int *result, int indice)
{
int temp;
int i, j;
for(i = 0; i < n; i++)
{
temp = 0;
for(j = 0; j < k; j++)
temp ^= vector[j] & matrix[j][i];
result[i + indice] = temp;
84
}
}
Phoenix Group
Oct-05
INRIA - LaBRI
length in = 4
length out = 7
type = "lin_b_code”
matrix = {{1, 1, 1}, {1, 1, 0},
{1, 0, 0}, {0, 1, 1}}
opt = "syst"
#include
#include
#include
#include
#include
#include
#include
"encode.h"
"parity_bit.h"
"lin_b_code.h"
"lin_b_code_sys.h"
"code_conv.h"
"encode_crc.h"
"libstring.h"
struct data
{
int longueur;
int *v;
};
void
encode(struct data *donnees, struct code *codage, struct data *result){
struct code
if (donnees->longueur != codage->k)
{
{ perror("donnees->longueur != codage->k");
int k;
exit(1);}
int n;
else if (strcmp(codage->type, "parity_bit") == 0)
int **matrix;
if (result->longueur != codage->k + 1)
char *type;
{ perror("result->longueur != codage->k + 1");
char *opt;
exit(1);}
};
else parity_bit(codage->k, donnees->v, result->v);
else if (strcmp(codage->type, "lin_b_code") == 0)
extern void encode(struct data *donnees, struct code *codage, struct data *result);
if ((*result).longueur != (*codage).n)
{ perror ("result->longueur != codage->n");
exit (1);}
else if (strcmp(codage->opt, "syst") == 0)
lin_b_code_sys(donnees->v, codage->k, codage->n, codage->matrix, result->v);
else
lin_b_code(donnees->v, codage->k, codage->n, codage->matrix, result->v);
else if (strcmp(codage->type, "code_conv") == 0)
if ((*result).longueur != (2 * (*codage).k))
{ perror ("result->longueur != (2 * codage->k)");
exit (1);}
else code_conv(donnees->v, codage->k, result->v);
else if (strcmp(codage->type, "crc") == 0)
if ((*result).longueur != (*codage).n )
{ perror ("result->longueur != codage->n");
exit (1); }
else encode_crc(donnees->v, codage->k, codage->n, codage->matrix[0], result->v);
else
{ perror("unknown code type");
extern int strcmp(char *s1, char *s2);
exit(1); }
}
extern void encode_crc(int *donnees, int k, int n, int *poly, int *result);
extern void parity_bit(int l, int *vector, int *result);
#include "parity_bit.h"
void
parity_bit(int l, int *vector, int *result)
{
int res;
int i;
res = 0;
for(i = 0; i < l; i++)
{
res ^= vector[i];
result[i] = vector[i];
}
result[l] = res;
extern void code_conv(int *donnees, int l, int *result);
extern void lin_b_code_sys(int *vector, int k, int n, int **matrix, int *result);
#include "code_conv.h"
void
code_conv(int *donnees, int k, int *result)
{
int temp;
int i;
#include "lin_b_code_sys.h"
#include "mult_mat.h"
void
lin_b_code_sys(int *vector, int k, int n, int **matrix, int *result)
{
int sub_matrix_size;
int i;
for (i = 0; i < k - 2; i++)
{
temp = donnees[i] ^ donnees[i + 2];
result[2 * i + 1] = temp;
temp ^= donnees[i + 1];
result[2 * i] = temp;
}
temp = donnees[k - 2] ^ donnees[k - 1];
result[2 * k - 4] = temp;
result[2 * k - 3] = donnees[k - 2];
result[2 * k - 2] = donnees[k - 1];
result[2 * k - 1] = donnees[k - 1];
sub_matrix_size = n - k;
mult_mat(vector, k, sub_matrix_size, matrix, result, k);
for(i = 0; i < k; i++)
result[i] = vector[i];
}
}
}
#include "encode_crc.h"
void
encode_crc(int *donnees, int k, int n, int *poly, int *result)
{
int l_poly;
int i, j;
l_poly = n - k + 1;
for (i = 0; i < k; i++)
result[i] = donnees[i];
for (i = 0; i < k; i++)
if (result[i] == 1)
for (j = 0; j < l_poly; j++)
result[i + j] ^= poly[j];
for (i = 0; i < k; i++)
result[i] = donnees[i];
}
extern void lin_b_code(int *vector, int k, int n, int **matrix, int *result);
#include "lin_b_code.h"
#include "mult_mat.h"
void
lin_b_code(int *vector, int k, int n, int **matrix, int *result)
{
mult_mat(vector, k, n, matrix, result, 0);
}
extern void mult_mat(int *vector, int k, int n, int **matrix, int *result, int indice);
#include "mult_mat.h"
void
mult_mat(int *vector, int k, int n, int **matrix, int *result, int indice)
{
int temp;
int i, j;
for(i = 0; i < n; i++)
{
temp = 0;
for(j = 0; j < k; j++)
temp ^= vector[j] & matrix[j][i];
result[i + indice] = temp;
}
85
}
Phoenix Group
Oct-05
INRIA - LaBRI
struct data { int longueur;
int *v; };
struct code { int k;
int n;
int **matrix;
char *type;
char *opt; };
extern int perror();
extern int exit();
extern void encode_spe(struct data *, struct data *);
length in = 4
length out = 7
type = "lin_b_code"
matrix = {{1, 1, 1}, {1, 1, 0},
{1, 0, 0}, {0, 1, 1}}
opt = "syst"
extern void encode_spe/*0*/(struct data *donnees, struct data *result)
{
{
int *lin_b_code_sys_0_result;
int *lin_b_code_sys_0_vector;
lin_b_code_sys_0_vector = (*donnees).v;
lin_b_code_sys_0_result = (*result).v;
{
int *mult_mat_1_result;
int *mult_mat_1_vector;
int mult_mat_1_temp;
mult_mat_1_vector = lin_b_code_sys_0_vector;
mult_mat_1_result = lin_b_code_sys_0_result;
mult_mat_1_temp = (((0 ^ (unsigned int)((int)((unsigned int)mult_mat_1_vector[0] & 1))) ^
(unsigned int)((int)((unsigned int)mult_mat_1_vector[1] & 1))) ^
(unsigned int)((int)((unsigned int)mult_mat_1_vector[2] & 1))) ^
(unsigned int)((int)((unsigned int)mult_mat_1_vector[3] & 0));
mult_mat_1_result[4] = mult_mat_1_temp;
mult_mat_1_temp = (((0 ^ (unsigned int)((int)((unsigned int)mult_mat_1_vector[0] & 1))) ^
(unsigned int)((int)((unsigned int)mult_mat_1_vector[1] & 1))) ^
(unsigned int)((int)((unsigned int)mult_mat_1_vector[2] & 0))) ^
(unsigned int)((int)((unsigned int)mult_mat_1_vector[3] & 1));
mult_mat_1_result[5] = mult_mat_1_temp;
mult_mat_1_temp = (((0 ^ (unsigned int)((int)((unsigned int)mult_mat_1_vector[0] & 1))) ^
(unsigned int)((int)((unsigned int)mult_mat_1_vector[1] & 0))) ^
(unsigned int)((int)((unsigned int)mult_mat_1_vector[2] & 0))) ^
(unsigned int)((int)((unsigned int)mult_mat_1_vector[3] & 1));
mult_mat_1_result[6] = mult_mat_1_temp;
}
lin_b_code_sys_0_result[0]
lin_b_code_sys_0_result[1]
lin_b_code_sys_0_result[2]
lin_b_code_sys_0_result[3]
=
=
=
=
lin_b_code_sys_0_vector[0];
lin_b_code_sys_0_vector[1];
lin_b_code_sys_0_vector[2];
lin_b_code_sys_0_vector[3];
}
return;
}
86
Phoenix Group
Oct-05
INRIA - LaBRI
Benchmarks
Handwritten encoder
Generic encoder
Customized encoder
87
Phoenix Group
Oct-05
INRIA - LaBRI
Conclusion


Customizable component based on a declarative approach
–
no hardcoded transformations
–
module aside from the code
–
customization properties
Environment
–
visualization tools
–
compiler for modules

code transformation

verification

Automatic generator of customized component

Applications: XDR library, packet filtering, FFT, system
interrupts…
88
Phoenix Group
Oct-05
INRIA - LaBRI
Specialization of Protocol Stacks in
OS Kernels
Part II
89
INRIA - LaBRI
Phoenix Group
Oct-05
Introduction


High-speed networks require efficient end-system
support
Efficient end-system support =
Efficient applications (clients, servers) +
Efficient protocol stacks
 For high throughput and low service latency, we need
efficient applications and protocol stacks
90
Phoenix Group
Oct-05
INRIA - LaBRI
Introduction


High-speed networks require efficient end-system
support
Efficient end-system support =
Efficient applications (clients, servers) +
Efficient protocol stacks
 For high throughput and low service latency, we need
efficient applications and protocol stacks
91
Phoenix Group
Oct-05
INRIA - LaBRI
Overview






92
Some data points
Application vs Protocol Stack optimization
Program Specialization
Specialization of the protocol stack
Performance Evaluation
Conclusion
Phoenix Group
Oct-05
INRIA - LaBRI
Some data points, to begin with

Physical transmission bound vs actual throughput
between two Pentium II-700Mhz machines with
100Mbps network (MTU 1500):
–
–
–

93
Max UDP throughput: (70 Mbps)
Max TCP throughput: (47 Mbps)
HTTP throughput at saturation: (28 Mbps)
All limited by the end system
Phoenix Group
Oct-05
INRIA - LaBRI
Application vs Protocol stack optimization
Application
Protocol Stack
GOAL
Exploit full b/w of
protocol stack
Exploit full b/w of
hardware
EXTENT
Limited to
application,
non-intrusive
Implemented in
kernel,
intrusive
DEPLOYED
Late
Early (With OS)
NATURE
Application-specific
“Application-specific”
94
Phoenix Group
Oct-05
INRIA - LaBRI
Application vs Protocol stack optimization
Application
GOAL
EXTENT
Best left to the
application
developer
Protocol Stack
Exploit full b/w of
hardware
Implemented in
kernel,
intrusive
Early (With OS)
DEPLOYED
“Application-specific”
NATURE
95
Phoenix Group
Oct-05
INRIA - LaBRI
Application vs Protocol stack optimization
Application
Protocol Stack
Best left to the
application
developer
Need a systematic
process to
customize…
GOAL
EXTENT
DEPLOYED
NATURE
96
Phoenix Group
Oct-05
INRIA - LaBRI
Program Specialization
GENERIC CODE
SPECIALIZED CODE
CONFIGURATION
VALUES
97
Program Specializer
(Tempo)
Phoenix Group
Oct-05
INRIA - LaBRI
Program Specialization
98
int tcp_mini_sendmsg (struct sock *sk, void *msg, int size)
{
int tocopy=0, copied=0;
while (tocopy = (size < sk->tcp->mss) ? size : mss) {
if (copied = (free_space (sk->write_queue.prev.space))) {
if (copied > tocopy) copied = tocopy;
add_data (sk->write_queue.prev, msg, copied);
size = size - copied; msg = msg + copied;
}
else {
struct skbuff *skb = alloc_new_skb();
add_data(skb, msg, tocopy);
size = size - tocopy; msg = msg + tocopy;
entail (sk->write_queue, skb);
}
}
return size;
}
Phoenix Group Oct-05
INRIA - LaBRI
Program Specialization
99
int tcp_mini_sendmsg (struct sock *sk, void *msg, int size)
{
int tocopy=0, copied=0;
while (tocopy = (size < sk->tcp->mss) ? size : mss) {
if (copied = (free_space (sk->write_queue.prev.space))) {
if (copied > tocopy) copied = tocopy;
add_data (sk->write_queue.prev, msg, copied);
size = size - copied; msg = msg + copied;
}
else {
struct skbuff *skb = alloc_new_skb();
add_data(skb, msg, tocopy);
size = size - tocopy; msg = msg + tocopy;
entail (sk->write_queue, skb);
}
}
return size;
}
Phoenix Group Oct-05
INRIA - LaBRI
Program Specialization
100
int tcp_mini_sendmsg (struct sock *sk, void *msg, int size)
{
int tocopy=0, copied=0;
while (tocopy = (size < sk->tcp->mss) ? size : mss) {
if (copied = (free_space (sk->write_queue.prev.space))) {
if (copied > tocopy) copied = tocopy;
add_data (sk->write_queue.prev, msg, copied);
size = size - copied; msg = msg + copied;
}
else {
struct skbuff *skb = alloc_new_skb();
add_data(skb, msg, tocopy);
size = size - tocopy; msg = msg + tocopy;
entail (sk->write_queue, skb);
}
}
return size;
}
Phoenix Group Oct-05
INRIA - LaBRI
Program Specialization
size=1400
sk={…}
Specialization context:
Specialized code:
101
int tcp_mini_sendmsg (void *msg)
{
struct skbuff *skb = alloc_new_skb();
add_data(skb, msg, 1400);
entail (sk->write_queue, skb);
return 0;
}
Phoenix Group
Oct-05
INRIA - LaBRI
How specialization is introduced
Application
OS
102
token = do_customize_send(…);
for (i=0;i<100000;i++) {
customized_send (token, buffer);
}
Phoenix Group
Oct-05
INRIA - LaBRI
How specialization is introduced
Application
OS
103
token = do_customize_send(…);
for (i=0;i<100000;i++) {
customized_send (token, buffer);
}
Phoenix Group
Oct-05
INRIA - LaBRI
How specialization is introduced
Application
OS
104
token = do_customize_send(…);
for (i=0;i<100000;i++) {
customized_send (token, buffer);
}
Phoenix Group
Oct-05
INRIA - LaBRI
Spec. opportunities in protocol stacks:
Eliminating Genericity





105
Eliminating lookups
Eliminating option interpretation
Eliminating routing decisions
Optimizing buffer allocation
Optimizing data fragmentation and coalescing
Phoenix Group
Oct-05
INRIA - LaBRI
Eliminating lookups
Socket
descriptors
106
Application
Phoenix Group
Oct-05
INRIA - LaBRI
Eliminating lookups
Socket
structures
Socket
descriptors
107
Protocol layer
Kernel
Application
Phoenix Group
Oct-05
INRIA - LaBRI
Eliminating lookups
Socket
structures
Protocol layer
Kernel
Socket i-node
Sockets interface
Socket
descriptors
108
Application
Phoenix Group
Oct-05
INRIA - LaBRI
Eliminating lookups
Kernel
Application
109
Phoenix Group
Oct-05
INRIA - LaBRI
Eliminating lookups
Kernel
sockfd_lookup(sd)
Application
110
Phoenix Group
Oct-05
INRIA - LaBRI
Eliminating socket options
BLOCKING
NAGLE
Application
UNICAST
111
Phoenix Group
Oct-05
INRIA - LaBRI
Eliminating socket options
UNICAST?
BLOCKING?
NAGLE?
BLOCKING
Kernel
NAGLE
Application
UNICAST
112
Phoenix Group
Oct-05
INRIA - LaBRI
Eliminating socket options
UNICAST?
BLOCKING?
NAGLE?
Kernel
sock_sendmsg(…, …, …, flags)
BLOCKING
NAGLE
Application
UNICAST
113
Phoenix Group
Oct-05
INRIA - LaBRI
Eliminating routing decisions
NIC driver
ip_route_output
Kernel
TCP/UDP layer
System call
114
Phoenix Group
Oct-05
INRIA - LaBRI
Eliminating routing decisions
NIC driver
ip_route_output
FIB
Kernel
TCP/UDP layer
System call
115
Phoenix Group
Oct-05
INRIA - LaBRI
Eliminating routing decisions
NIC driver
ip_route_output
Route cache
Kernel
TCP/UDP layer
System call
116
Phoenix Group
Oct-05
INRIA - LaBRI
Eliminating routing decisions
NIC driver
ip_route_output
Kernel
TCP/UDP layer
System call
connect (int sockfd, struct sockaddr *daddr, int addrlen)
117
Phoenix Group
Oct-05
INRIA - LaBRI
Optimizing buffer allocation
if (in interrupt() && (gfp mask & GFP WAIT)) {
gfp mask &= ~GFP WAIT;
npages = (dlen + (PAGE SIZE- 1))
>> PAGE SHIFT;
skb->truesize += dlen;
((struct skb sharedinfo *)
skb->end)->nr frags = npages;
for (i = 0; i < npages; i++) {
...
}
}
118
Phoenix Group
Oct-05
Kernel
INRIA - LaBRI
Optimizing buffer allocation
if (in interrupt() && (gfp mask & GFP WAIT)) {
gfp mask &= ~GFP WAIT;
npages = (dlen + (PAGE SIZE- 1))
>> PAGE SHIFT;
skb->truesize += dlen;
((struct skb sharedinfo *)
skb->end)->nr frags = npages;
for (i = 0; i < npages; i++) {
...
}
}
Kernel
setsockopt(…, …, …, SO_FIXADU, &size, …)
119
Phoenix Group
Oct-05
INRIA - LaBRI
Optimizing coalescing and fragmentation
while (seglen > 0)
{
copy = MSS_now – last_skb_len;
if (copy > 0) {
if (copy < seglen)
copy = seglen;
push_into_previous(copy);
}
else {
copy = min(seglen, MSS_now);
push_into_current(copy);
}
seglen -= copy;
}
Kernel
setsockopt(…, …, …, SO_FIXADU, &size, …)
120
Phoenix Group
Oct-05
INRIA - LaBRI
Specifying the opportunities to the code
specializer
Original C code:
struct sk_buff *sock_alloc_send_pskb( struct sock *sk,
unsigned long header len,
unsigned long data len,
int noblock,
int *errcode) {
...
}
121
Specialization declarations:
Sock alloc_send_pskb:: intern sock_alloc_send_pskb(
Spec sock( struct sock) S(*) sk,
S( unsigned long ) header len,
S( unsigned long ) data len,
S( int ) noblock,
D( int * ) errcode) {
...
};
Phoenix Group Oct-05
INRIA - LaBRI
Results: Improvements in
Performance and Code Size



Execution time decreased by ~25%
Code size decreased by a factor of >15
Throughput improvements:
–
–
122
UDP - PIII: 13% 486: 27% iPAQ: 18%
TCP - PIII: 10% 486: 23% iPAQ: 13%
Phoenix Group
Oct-05
INRIA - LaBRI
Specialization Overhead

Specialization
–
–

Bottleneck
–
–

123
Run time for fast specialization
Compile time for fast specialized code
Execution of specialization
Compiler
What about performing specialization remotely?
Phoenix Group
Oct-05
INRIA - LaBRI
Conclusion



124
Problem: Need for fast protocol stacks in high
speed networks
Solution: Program specialization at run time
Assessment: Execution time up to 25%, code size
up to 15x, throughput up to 20%
Phoenix Group
Oct-05
INRIA - LaBRI
Remote Customization Of Systems Code
For Embedded Devices
Part III
12
5
INRIA - LaBRI
Phoenix Group
Oct-05
Problem Statement


Embedded systems low on resources
OSes generic
Overheads (space, time)
Need for customization of systems code
126
Phoenix Group
Oct-05
INRIA - LaBRI
Outline







127
Introduction
Code Customization
Remote Customization Infrastructure
Virtualization of memory
Case study: TCP/IP
Performance Evaluation
Conclusion
Phoenix Group
Oct-05
INRIA - LaBRI
Outline






128
Introduction
Remote Customization Infrastructure
Virtualization of memory
Case study: TCP/IP
Performance Evaluation
Conclusion
Phoenix Group
Oct-05
INRIA - LaBRI
Dedicated Vs. Generic OSes
Generic OSes
Dedicated OSes
+ Deeply customized, compact,
fast, well-suited
- Lack of support for standards
129
Phoenix Group
+ Support for standards
- Generic, coarse grained
abstractions
Oct-05
INRIA - LaBRI
Industry Trends
Percent
Embedded Systems Developers Survey, 2002
Source: Evans Data Corp.
130
Phoenix Group
Oct-05
INRIA - LaBRI
Generic abstractions
Coarse grained building
blocks
if (poll (listen_pfds, n, -1) > 1) {
foreach(pfd, listen_pfds) {
if (hi_r(pfd->revents))
queue(
accept(fd, addr, addr_len));
}
}
Performance:
3000+ conn/sec.
Concrete Operations
foreach(sock, my_sockets) {
if (sock->sk->accept_queue) {
sock->ops->accept(sock, new_sock,
O_NONBLOCK);
}
}
Performance:
7000+ conn/sec.
Overheads: memory transfers, context switches,
sanity checks, data structures
131
Phoenix Group
Oct-05
INRIA - LaBRI
Generic abstractions
Coarse grained building
blocks
if (poll (listen_pfds, n, -1) > 1) {
foreach(pfd, listen_pfds) {
if (hi_r(pfd->revents))
queue(
accept(fd, addr, addr_len));
}
}
Performance:
3000+ conn/sec.
Concrete Operations
foreach(sock, my_sockets) {
if (sock->sk->accept_queue) {
sock->ops->accept(sock, new_sock,
O_NONBLOCK);
}
}
Performance:
7000+ conn/sec.
Need for Program Specialization
132
Phoenix Group
Oct-05
INRIA - LaBRI
Outline






133
Introduction
Remote Customization Infrastructure
Virtualization of memory
Case study: TCP/IP
Performance Evaluation
Conclusion
Phoenix Group
Oct-05
INRIA - LaBRI
Remote Customization
Customized
Code
Remote customization
server
134
Network
Phoenix Group
Oct-05
INRIA - LaBRI
Specialization Opportunity
Application
for (i=0 ; i < 100000; i++) {
send(…, buffer);
}
OS
135
Phoenix Group
Oct-05
INRIA - LaBRI
How it’s used
Application
OS
136
token = do_customize_send(…);
for (i=0 ; i < 100000; i++) {
customize_send(token, buffer);
}
Phoenix Group
Oct-05
INRIA - LaBRI
How it’s used
Application
OS
137
token = do_customize_send(…);
for (i=0 ; i < 100000; i++) {
customize_send(token, buffer);
}
Phoenix Group
Oct-05
INRIA - LaBRI
How it’s used
Application
OS
138
token = do_customize_send(…);
for (i=0 ; i < 100000; i++) {
customize_send(token, buffer);
}
Phoenix Group
Oct-05
INRIA - LaBRI
Architecture
User space
Customizer
Compiler
Application
Runtime Layer
Context
Manager
139
Kernel Space
Code
Manager
Phoenix Group
Context
Manager
Oct-05
Code
Manager
INRIA - LaBRI
Customization Request
Customizer
Compiler
Runtime Layer
Code
Manager
Context
Manager
User space
Customization
Request
Application
Kernel Space
Context
Manager
Code
Manager
Application issues customization request
140
Phoenix Group
Oct-05
INRIA - LaBRI
Context Manager
Customizer
Compiler
Runtime Layer
Code
Manager
Context
Manager
User space
Application
syscall=sys_send
fd=4;
daddr=1044321;
flags=32;
...
...
Kernel Space
Context
Manager
Code
Manager
Context manager picks up customization context
141
Phoenix Group
Oct-05
INRIA - LaBRI
Code Manager
Customizer
Compiler
Runtime Layer
Code
Manager
Context
Manager
User space
Application
Kernel Space
Context
Manager
Code
Manager
Check if we have code for the current context
142
Phoenix Group
Oct-05
INRIA - LaBRI
Customization Request
Customizer
fd=4;
daddr=1044321;
flags=32;
addr_len=8;
block_size=1483;
Compiler
(...)
Runtime Layer
Context
Manager
User space
Application
Kernel Space
Code
Manager
Context
Manager
Code
Manager
Application issues customization request
143
Phoenix Group
Oct-05
INRIA - LaBRI
Runtime Layer
User space
Customizer
Compiler
Application
Runtime Layer
Context
Manager
Kernel Space
Code
Manager
Context
Manager
Code
Manager
Context manager invokes runtime layer
144
Phoenix Group
Oct-05
INRIA - LaBRI
Customizer
User space
Customizer
Compiler
Application
Runtime Layer
Context
Manager
Kernel Space
Code
Manager
Context
Manager
Code
Manager
The program customizer is Tempo
145
Phoenix Group
Oct-05
INRIA - LaBRI
Compiler
User space
Customizer
Compiler
Application
Runtime Layer
Context
Manager
Kernel Space
Code
Manager
Context
Manager
Code
Manager
The customized code is compiled using a standard compiler
146
Phoenix Group
Oct-05
INRIA - LaBRI
Code Manager
User space
Customizer
Compiler
Application
Runtime Layer
Context
Manager
Kernel Space
Code
Manager
Context
Manager
Code
Manager
Customized code is sent back
147
Phoenix Group
Oct-05
INRIA - LaBRI
Customization Token
User space
Customizer
Compiler
Application
Runtime Layer
Context
Manager
148
Kernel Space
Code
Manager
Phoenix Group
Context
Manager
Oct-05
Code
Manager
INRIA - LaBRI
Customization Token
Customizer
Compiler
Runtime Layer
Code
Manager
Context
Manager
User space
Application
Customization
Token
(eg., 0 for the
first
customization)
Kernel Space
Context
Manager
Code
Manager
Application gets back a customization token
149
Phoenix Group
Oct-05
INRIA - LaBRI
Customization Syscall
Customizer
Compiler
Runtime Layer
Code
Manager
Context
Manager
User space
Per-process syscall
table
Application
Kernel Space
Context
Manager
Code
Manager
Application uses customization token as an index
150
Phoenix Group
Oct-05
INRIA - LaBRI
Outline







151
Introduction
Code Customization
Remote Customization Infrastructure
Virtualization of memory
Case study: TCP/IP
Performance Evaluation
Conclusion
Phoenix Group
Oct-05
INRIA - LaBRI
Access to client side memory
movl [socket_pointer],%eax
Customizer
0xc01f400: 1483 [tcp_mss]
0xc01f363: 0xc01f355 [tp]
Runtime Layer
152
1. Run-time layer intercepts
dereference, as CPU exception.
2. Run-time layer interprets
instruction with values in
customization context table.
Phoenix Group
Oct-05
INRIA - LaBRI
Access to client side memory
movl [socket_pointer],%eax
Customizer
0xc01f400: 1483 [tcp_mss]
0xc01f363: 0xc01f355 [tp]
Runtime Layer
153
1. Run-time layer intercepts
dereference, as CPU exception.
2. Run-time layer interprets
instruction with values in
customization context table.
3. Customization-time functions
Executed on client
Phoenix Group
Oct-05
INRIA - LaBRI
Outline







154
Introduction
Code Customization
Remote Customization Infrastructure
Virtualization of memory
Case study: TCP/IP (already presented)
Performance Evaluation
Conclusion
Phoenix Group
Oct-05
INRIA - LaBRI
Outline







155
Introduction
Code Customization
Remote Customization Infrastructure
Virtualization of memory
Case study: TCP/IP
Performance Evaluation (already presented)
Conclusion
Phoenix Group
Oct-05
INRIA - LaBRI
Conclusion

Problem:
–

Solution:
–
–

156
Services in generic OSes are slow and bloated
Dynamic program specialization
Remote specialization  limited resources
Assessment: Exec time… -25%, throughput…
+20%, code size… -15x
Phoenix Group
Oct-05
INRIA - LaBRI