Profile-Driven Selective Program Loading Tugrul Ince tugrul@cs.umd.edu Jeff Hollingsworth Department of Computer Science University of Maryland, College Park, MD 20742 University of Maryland Motivation Programs are getting larger! – Many frameworks and libraries Many supercomputers lack demand-paging – Example: Cray XT and BlueGene series – Available memory is scarce Observation: Most programs do not use every available function! – Frameworks and libraries are too general – Code that handles errors or special cases 2 Why not remove functions that are not used in the common case? University of Maryland Aim Reduce memory footprint by selectively loading parts of shared libraries 3 University of Maryland Target Platforms and Applications Unix/Linux systems that support ELF – Modifies ELF program headers Applications with many libraries – Most current reasonable applications Parallel programs running on multiple nodes – MPI etc. Platforms without demand-paging – Cray XT and BlueGene series 4 University of Maryland Architecture Overview Application is profiled. It is rewritten with – Modified Shared Libraries – A Signal Handler 5 Application is executed as usual. University of Maryland Profiler Need a list of never-called functions in each shared library – Profile the application several times – May not be perfect DynInst-based profiler – Write small program (~ 70 LOC) – Rewrite shared libraries – Profile as many times as necessary 6 University of Maryland Rewriting Do not load unused functions .text – Modify ELF program headers – Example: libpetsc.so Program Headers: Type LOAD LOAD DYNAMIC GNU_STACK Offset 0x000000 0x124584 0x112000 0x12459c 0x000000 VirtAddr 0x00000000 0x00125584 0x00112000 0x0012559c 0x00000000 PhysAddr 0x00000000 0x00125584 0x00112000 0x0012559c 0x00000000 FileSiz 0x090000 0x124584 0x013f8 0x012584 0x00130 0x00000 First Loadable Section: .text, .init, .fini, .plt 7 Second Loadable Section: .dynamic, .got, .got.plt, .data, .bss University of Maryland MemSiz 0x090000 0x124584 0x0a434 0x012584 0x00130 0x00000 Flg R E RWE R RW RW Align 0x1000 0x1000 0x4 0x4 Rewriting Do not load unused functions .text – Modify ELF program headers – Example: libpetsc.so Program Headers: Type LOAD LOAD LOAD DYNAMIC GNU_STACK Offset 0x000000 0x112000 0x124584 0x12459c 0x000000 VirtAddr 0x00000000 0x00112000 0x00125584 0x0012559c 0x00000000 PhysAddr 0x00000000 0x00112000 0x00125584 0x0012559c 0x00000000 FileSiz 0x090000 0x012584 0x013f8 0x00130 0x00000 First Loadable Section: .text, .init, .fini, .plt 8 Second Loadable Section: .dynamic, .got, .got.plt, .data, .bss University of Maryland MemSiz 0x090000 0x012584 0x0a434 0x00130 0x00000 Flg R E R E RW RW RW Align 0x1000 0x1000 0x1000 0x4 0x4 Rewriting Rewriter based on DynInst Profile data is used to create lists of Used and Unused functions Access / Modify symbols Defragment functions to maximize space savings – Requires moving functions inside shared libraries 9 University of Maryland Function Defragmentation Used Unused 10 University of Maryland Challenges: Relative Calls Common way of calling functions in PIC. If either callee or caller is moved, their relative positioning changes. Offsets in such relative call instructions need to be updated call d call d’ d d' foo foo 11 University of Maryland Challenges: Symbols Runtime linker uses symbols to resolve cross-library calls. – Uses procedure linkage tables (plt) If a function is moved, its associated symbol has to be updated. foo: 0xdeadbeef foo@plt foo: 0xbeefdead foo@plt foo call foo@plt 12 call foo@plt foo University of Maryland Challenges: Jump Tables Used to represent n-way branches at machine level Targets are read from jump table – Entries are offsets of targets from the GOT address 13 Becomes invalid if the function referenced in a jump table is moved DynInst reads jump tables to generate CFGs We update entries so that they can be used to point to new location of targets University of Maryland Unexpectedly Called Function Execution is not always predictable – Unexpected function calls Rewrite original executable with a Signal Handler Load the function upon an unexpected call – Signal Handler picks up page faults (SIGSEGV) – Loads requested page on-demand – Execution resumes 14 User-level: No OS modifications University of Maryland Experiments Tested on – PETSc ex5 in snes package – PETSc ex2 in ksp package – GS2 Compiled with debug flag and no optimization Used Open MPI Tested on 64-node cluster at UMD – Dual-core x86 processors – Unmodified Linux kernel 15 Space savings of about 82% on average University of Maryland PETSc – snes (ex5) Library Name Text Pages (Original) Text Pages (Modified) Reduction % petsc 260 68 73.85 petscdm 161 19 88.2 petscksp 335 39 88.36 blas petscmat 772 40 94.82 petscvec 204 52 74.51 petscsnes 20 20 0 mpi_cxx 10 5 50 142 37 73.94 gfortran open-pal 62 34 45.16 open-rte 55 34 m 28 3 mpi Library Name Text Pages (Original) Text Pages (Modified) Reductio n% X11 146 7 95.21 lapack 866 2 99.77 80 3 96.25 stdc++ 133 12 90.98 gcc_s 12 2 83.33 Xau 2 2 0 Xdcm 3 3 0 123 4 96.75 dl 2 2 0 38.18 nsl 14 2 85.71 89.29 util 2 2 0 2021 348 OVERALL 16 University of Maryland 82.78 PETSc – snes (ex5) (2021) 1000 900 800 700 600 500 Original Modified 400 300 200 100 0 17 University of Maryland PETSc – ksp (ex2) Library Name Text Pages (Modified) Reduction % petsc 260 72 72.31 petscdm 161 3 98.14 petscksp 335 49 85.37 petscmat 772 49 93.65 petscvec 204 54 73.53 mpi_cxx 10 5 50 142 47 66.9 open-pal 62 37 40.32 open-rte 55 36 34.55 OVERALL 2001 352 82.41 mpi 18 Text Pages (Original) University of Maryland GS2 Library Name Reduction % MdsLib 21 0 100 MdsShr 21 0 100 TdiShr 220 3 98.64 TreeShr 38 0 100 fftw 70 25 64.29 rfftw 58 8 86.21 mpi_f77 13 2 84.62 142 40 71.83 open-pal 62 36 41.94 open-rte 55 36 34.55 OVERALL 700 150 78.57 mpi 19 Text Pages (Original) Text Pages (Modified) University of Maryland Running Times GS2 takes 5 seconds less on average – (36m 38s vs. 36m 33s) Overhead on PETSc examples – ex2 runs for 2.7 secs, ex5 runs for 1.05 secs. 20 University of Maryland Running Times Results suggest no overhead for reasonably-long running programs – Initial cost for signal handler registration – Better instruction cache and TLB performance 21 University of Maryland Summary Our tool reduces memory footprint of shared libraries Rewrite shared libraries with holes – Defragment functions to maximize space savings 22 On-demand page loading if a not-yetloaded function is called About 82% memory space savings for shared libraries Might improve instruction cache and TLB performance University of Maryland