GHUMVEE: Efficient, Effective and Flexible Replication Stijn - PowerPoint PPT Presentation

Vakgroep ELIS GHUMVEE: Efficient, Effective and Flexible Replication Stijn Volckaert Computer Systems Lab Ghent University Belgium

N-modular Redundancy variant 1 input variant 2 monitor output variant 3 equivalent components 2

Replication Variant 1 Variant 2 libc libc t t sys_read sys_read sys_exec sys_read sys_read sys_read Monitor Kernel Hardware 3

Replication Goals  Run variants in parallel on the same inputs  Detect inconsistent behavior  Transparent to user and programmer  Minimal overhead  Support wide range of diversity  Run realistic programs Cox, B., Evans, D., et al.: N-variant systems: A secretless framework for security through diversity. In: Proc. USENIX SSYM. (2006) 105-120 • Berger, E., Zorn, B.: DieHard: probabilistic memory safety for unsafe languages. In: Proc. ACM PLDI. (2006) 158-168 • Bruschi, D., Cavallaro, L.: Diversifed Process Replicae for Defeating Memory Error Exploits. In: Proc. IEEE IPCCC. (2007) 434-441 • Salamat, B., Jackson, T., et al.: Orchestra: A User Space Multi-Variant Execution Environment. In: Proc. EuroSys. (2009) 33-46 • 4

Overview  Introduction  Replication  GHUMVEE Overview  Implementation challenges  Multithreading & synchronization  Address-sensitive behavior  Evaluation  Conclusions 5

Multithreading (1) Variant1 Variant1 Variant2 Variant2 Variant 1 Variant 2 Thread1 Thread2 Thread1 Thread2 t t t t sys_brk sys_open Kernel Hardware 6

Multithreading (2) Variant1 Variant1 Variant2 Variant2 Thread1 Thread2 Thread1 Thread2 t t t t sys_open sys_open sys_brk sys_brk Kernel Hardware 7

User-space locking operations (1) Variant1 Variant2 Variant1 Variant2 Thread1 Thread1 Thread2 Thread2 t t t t sys_open sys_brk Kernel Hardware 8

User-space locking operations (2) Variant1 Variant2 Variant1 Variant2 1 Thread1 Thread1 Thread2 Thread2 t t t t sys_open sys_open Kernel Hardware 9

Address-Sensitive Behavior Variant 1 Variant 2 Object 3 (0x7c756c) Object 1 (0xd4cab9) Object 2 (0xdcd4c7) Object 1 (0xf0ebe2) Object 2 (0xb8a98f) sys_mmap2 Kernel 10

Benchmarks SPEC2006 Benchmarks 100 90 Relative Performance 80 70 60 50 40 30 20 10 0 no MVEE (1 variant) no MVEE (2 variants) MVEE (2 variants) MVEE (3 variants) MVEE (4 variants) 11 Measured on a core i7-870 quad core system

Supported Programs 12

Problematic features Multi- Custom Address Shared Mem- Time- Self- threade Sync Sensitive Mem mapped I/ Aware aware (/ d O (rdtsc) proc) Glibc Glib (GNOME) kcalc firefox LibreOffice MPlayer

Transparency  For the user:  Startup overhead  For the programmers:  Indicate names of functions that need interception  Don’t inline these functions interposer library interposer standard library libc pthread total (header files) base lib lines of C code 260 654 766 829 2509 application library glib gtk orbit pango libreoffice total lines of C code 105 54 78 54 183 474 14

Conclusions  Realistic programs  Limited performance overhead (~15%)  Limitations for programmers 15

16