情報処理学会研究報告 IPSJ SIG Technical Report Vol.2011-OS-118 No /7/28 LLVM LLVM Scattaring Object files by LLVM Natsuki Kawai 1 and Koichi Sa

Size: px

Start display at page:

Download "情報処理学会研究報告 IPSJ SIG Technical Report Vol.2011-OS-118 No /7/28 LLVM LLVM Scattaring Object files by LLVM Natsuki Kawai 1 and Koichi Sa"

るるみなつ
4 years ago
Views:

LLVM 1 1 1 1 1 LLVM Scattaring Object files by LLVM Natsuki Kawai 1 and Koichi Sasada 1 This paper describes the system scatters executable files or shared libraries, generated by compile and link

On the other hand, we can controll the loading process of a scattered program more flexibly and effectively.

1 LLVM LLVM Scattaring Object files by LLVM Natsuki Kawai 1 and Koichi Sasada 1 This paper describes the system scatters executable files or shared libraries, generated by compile and link processes, into a small launch program and numbers of shared libraries. When we execute a program, we load executable file from a storage to a memory. On the other hand, we can controll the loading process of a scattered program more flexibly and effectively. Our system assumes that a function is a smallest part of an object file and scatters an object file into libraries each of them contains one function. Our system also generates a launch program loads and executes the scattered shared libraries. To avoid a size-overhead of each shared library, our system generates function groups. We implemented our system on LLVM, a commonly used compiler infrastructure, to target many programs. As one of the applications of this system, this paper describes a way to run programs over a narrow network. Moreover, this paper evaluates file sizes and performance of programs scatterd by our system Fig. 1 Scattering a Program OS LLVM 1) Graduate School of Information Science and Technology The University of Tokyo 1 C 1 c 2011 Information

2 LLVM C Fig. 2 Functionality of the Scattering System c 2011 Information

3 ID ID 1 ID 3 Fig. 3 Generating a Shared Library subst ( ID 16 ) 3 c 2011 Information

4 movl $1 %eax jmp shared-lib manager 5 x86+cdecl Fig. 5 A Stub under x86+cdecl Environment 4 Fig. 4 Embed Stubs 1 4 ID ID ID CPU ID 2 x86 cdecl movl jmp 2 10byte ID 1 $( ID 16 ) 6 Fig. 6 Shared Library Manager 6 2 ID C main 2 NULL ID c 2011 Information

2 LLVM Pass Framework Pass Framework LLVM IR 7 1 1 LLVM IR 4.3 IR C global value 1 7 LLVM IR Fig.

5 ID NULL ID ID dlopen ID dlopen dlsym ID 1 4. LLVM LLVM 4.1 LLVM LLVM 1) LLVM LLVM LLVM Pass Framework 2) Pass Framework LLVM LLVM IR 4.2 LLVM Pass Framework Pass Framework LLVM IR LLVM IR 4.3 IR C global value 1 7 LLVM IR Fig. 7 Generating LLVM IR C global value IR 3) external common internal private 4 external common C internal private 3) internal private 3.1 internal private external LLVM IR Structure Structure Structure 5 c 2011 Information

1 1 1 1 1 5.1 5.2 8 LLVM IR Fig. 8 Scattering LLVM IR 5.

6 Structure ( 1 ) Structure ( 2 ) Structure ( 3 ) Structure 4.5 ID ID LLVM IR ID 4.6 LLVM IR 1 1 LLVM IR 1 1 IR 1 IR IR LLVM IR Fig. 8 Scattering LLVM IR 5.1 ( 1 ) G ( 2 ) 1 F ( a ) G F n F G ( b ) ( i ) G 6 c 2011 Information

7 ( ii ) G ( iii ) F G ( 3 ) 2 a h 4KiB, 8KiB, 6KiB, 48KiB, 4KiB, 4KiB, 4KiB, 4KiB 16KiB a, b 0 c 1 1 c d 2 e, f, g, h 3 ID ID 0 ID LLVM IR IR ID g( ID 16 ).so 5.2 ID ID ID ID dlopen ID NULL ID ID NULL ID ID dlsym NULL ID dlopen 6. HTTP SSH 1 NFS Linux ELF LOAD 7 c 2011 Information

8 NFS mmap 6.2 ID dlopen pthread dlopen dlsym ID dlopen dlopen dlopen 7. 1 Table 1 Environment OS Fedora 14 32bit (Linux ) CPU Intel Core2 Duo E6850 (3.00GHz) 3.2GiB Clang 1.1 LLVM 2.7 ] strip Vim-init Ruby-init Ruby-tarai SQLite3-init SQLite3-query 2 Table 2 Target Program Vim 7.3 Ruby p180 SQLite SQLite3 libsqlite3.so 2 demo1 9 demo c 2011 Information

9 Table 3 3 Each Function to a Shared Library (Launch Program Size) void func0000(void) { /* some calculations */ } void func0001(void) { /* some calculations */ } /* Total functions... */ int main(void) { } func0000(); func0400(); func0800(); /* Total 64 function calls... */ 9 1 Fig. 9 Sample Program 1 Table 4 [KiB] [KiB] Vim 417 2, % Ruby 479 2, % SQLite % demo1 3,205 1, % demo % 4 Each Function to a Shared Library (Total Shared Library Size) [KiB] [KiB] Vim 3,984 16,193 2, % Ruby 4,725 15,086 2, % SQLite3 3,604 11, % demo1 65, ,354 12, % demo % int main(void) { } volatile unsigned i, j, k; for (i = 0; i ; i++) for (j = 0; j 10000; j++) /* Null statement */; 10 2 Fig. 10 Sample Program % demo ms c 2011 Information

10 Table 5 5 Each Function to a Shared Library (Performance) Table 7 7 A Function Group to a Shared Library (Total Shared Library Size) Table 6 [ms] [ms] Ruby-init % Ruby-tarai 3,708 3, % Vim-init 2,289 2, % SQLite3-init % SQLite3-query 34,575 34, % demo % demo2 23,927 23, % 6 A Function Group to a Shared Library (Launch Program Size) [KiB] [KiB] Vim % Ruby % SQLite % demo1 3,206 3, % demo % KiB Linux [KiB] [KiB] [KiB] , % Vim ,309 2, % , % , % Ruby ,587 2, % , % , % SQLite , % , % , % demo ,914 12, % , % % demo % % SSHFS demo1 Ruby Vim 8 demo2 1Mbps SQLite3-query 8. 5) LLVM 10 c 2011 Information

11 11 LLVM 8 Table 8 Over a Narrow Bandwidth Network [kbps] [s] [s] Vim-init Ruby-init Rubytarai SQLite3- init SQLite3- query demo1 demo % % % % % % % % % % % % % % % % % % % % % 4) OS 6) 7) 8) SSHFS 60% NFS 11 c 2011 Information

12 1) Chris Lattner, Vikram Adve: LLVM: A Compilation Framework for Lifelong Program Analysis & Transformation, Proceedings of the international symposium on Code generation and optimization: feedback-directed and runtime optimization,p.75, March 20-24, 2004, Palo Alto, California 2) Chris Lattner, Jim Laskey: Writing an LLVM Pass, 3) Chris Lattner, Vikram Adve: LLVM Language Reference Manual, 4) Doug Joseph, Dirk Grunwald: Prefetching using Markov Predictors, ISCA 97 Proceedings of the 24th annual international symposium on Computer architecture 5) Saumya Debray, William Evans: Profile-Guided Code Compression, PLDI 02 Proceedings of the ACM SIGPLAN 2002 Conference on Programming language design and implementation 6) Rajeev Balasubramonian, David Albonesi, Alper Buyuktosunoglu, Sandhya Dwarkadas: Memory Hierarchy Reconfiguration for Energy and Performance in General-Purpose Processor Architectures, Proceedings of the 33rd annual ACM/IEEE international symposium on Microarchitecture (2000) 7) Ashutosh S. Dhodapkar, James E. Smith: Managing Multi-Configuration Hardware via Dynamic Working Set Analysis, ISCA 02 Proceedings of the 29th annual international symposium on Computer architecture 8) Doron Nakar, Shlomo Weiss: Selective Main Memory Compression by Identifying Program Phase Changes, WMPI 04 Proceedings of the 3rd workshop on Memory performance issues: in conjunction with the 31st international symposium on computer architecture 12 c 2011 Information

28 Docker Design and Implementation of Program Evaluation System Using Docker Virtualized Environment

28 Docker Design and Implementation of Program Evaluation System Using Docker Virtualized Environment 1170288 2017 2 28 Docker,.,,.,,.,,.,. Docker.,..,., Web, Web.,.,.,, CPU,,. i ., OS..,, OS, VirtualBox,.,