[BUG] Crash running parquet reader benchmarks.

[nvdbaranec]

The PARQUET_READER_NVBENCH crashes (segfault) at exit on some machines. It doesn't seem to happen consistently for everyone, but it tends to be reproducible once it starts happening.

To reproduce, run PARQUET_READER_NVBENCH and you should get a segfault right at the end after it has printed out all of it's results.

I've narrowed it down to something specific to the parquet_read_io_compression suite. In addition, compute-sanitizer does not turn anything up so this seems to be something purely cpu-side.

[wence-]

The first benchmark there doesn't appear to be valgrind-clean which may give a hint:

==45341== Warning: set address range perms: large range [0x300200000, 0x8f41ff000) (noaccess)
Run:  [1/24] parquet_read_io_compression [Device=0 io=FILEPATH compression=SNAPPY cardinality=0 run_length=1]
==45341== Invalid read of size 4
==45341==    at 0x7CC2078: cudf::io::detail::parquet::writer::impl::write(cudf::table_view const&, std::vector<cudf::io::partition_info, std::allocator<cudf::io::partition_info> > const&) (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/libcudf.so)
==45341==    by 0x7CC56E0: cudf::io::detail::parquet::writer::write(cudf::table_view const&, std::vector<cudf::io::partition_info, std::allocator<cudf::io::partition_info> > const&) (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/libcudf.so)
==45341==    by 0x7B9F8FA: cudf::io::write_parquet(cudf::io::parquet_writer_options const&) (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/libcudf.so)
==45341==    by 0x1588E1: parquet_read_common(cudf::io::parquet_writer_options const&, cuio_source_sink_pair&, nvbench::state&) (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/benchmarks/PARQUET_READER_NVBENCH)
==45341==    by 0x15F8D5: void BM_parquet_read_io_compression<(cudf::io::io_type)0, (cudf::io::compression_type)2>(nvbench::state&, nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)0, cudf::io::io_type>, nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type> >) [clone .isra.0] (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/benchmarks/PARQUET_READER_NVBENCH)
==45341==    by 0x15FACE: void nvbench::tl::detail::foreach<nvbench::type_list<nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)0, cudf::io::io_type>, nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)0, cudf::io::io_type>, nvbench::enum_type<(cudf::io::compression_type)0, cudf::io::compression_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)1, cudf::io::io_type>, nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)1, cudf::io::io_type>, nvbench::enum_type<(cudf::io::compression_type)0, cudf::io::compression_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)2, cudf::io::io_type>, nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)2, cudf::io::io_type>, nvbench::enum_type<(cudf::io::compression_type)0, cudf::io::compression_type> > >, nvbench::runner<nvbench::benchmark<BM_parquet_read_io_compression_line_142, nvbench::type_list<nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)0, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)1, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)2, cudf::io::io_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type>, nvbench::enum_type<(cudf::io::compression_type)0, cudf::io::compression_type> > > > >::run_device(std::optional<nvbench::device_info> const&)::{lambda(auto:1)#1}, 0ul, 1ul, 2ul, 3ul, 4ul, 5ul>(std::integer_sequence<unsigned long, 0ul, 1ul, 2ul, 3ul, 4ul, 5ul>, nvbench::runner<nvbench::benchmark<BM_parquet_read_io_compression_line_142, nvbench::type_list<nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)0, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)1, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)2, cudf::io::io_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type>, nvbench::enum_type<(cudf::io::compression_type)0, cudf::io::compression_type> > > > >::run_device(std::optional<nvbench::device_info> const&)::{lambda(auto:1)#1}&&) [clone .isra.0] (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/benchmarks/PARQUET_READER_NVBENCH)
==45341==    by 0x161D09: nvbench::runner<nvbench::benchmark<BM_parquet_read_io_compression_line_142, nvbench::type_list<nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)0, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)1, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)2, cudf::io::io_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type>, nvbench::enum_type<(cudf::io::compression_type)0, cudf::io::compression_type> > > > >::run() (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/benchmarks/PARQUET_READER_NVBENCH)
==45341==    by 0x161DAC: nvbench::benchmark<BM_parquet_read_io_compression_line_142, nvbench::type_list<nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)0, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)1, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)2, cudf::io::io_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type>, nvbench::enum_type<(cudf::io::compression_type)0, cudf::io::compression_type> > > >::do_run() (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/benchmarks/PARQUET_READER_NVBENCH)
==45341==    by 0x12837E: main (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/benchmarks/PARQUET_READER_NVBENCH)
==45341==  Address 0x83739940 is 0 bytes after a block of size 272 alloc'd
==45341==    at 0x4849013: operator new(unsigned long) (in /usr/libexec/valgrind/vgpreload_memcheck-amd64-linux.so)
==45341==    by 0x7CC1665: cudf::io::detail::parquet::writer::impl::write(cudf::table_view const&, std::vector<cudf::io::partition_info, std::allocator<cudf::io::partition_info> > const&) (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/libcudf.so)
==45341==    by 0x7CC56E0: cudf::io::detail::parquet::writer::write(cudf::table_view const&, std::vector<cudf::io::partition_info, std::allocator<cudf::io::partition_info> > const&) (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/libcudf.so)
==45341==    by 0x7B9F8FA: cudf::io::write_parquet(cudf::io::parquet_writer_options const&) (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/libcudf.so)
==45341==    by 0x1588E1: parquet_read_common(cudf::io::parquet_writer_options const&, cuio_source_sink_pair&, nvbench::state&) (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/benchmarks/PARQUET_READER_NVBENCH)
==45341==    by 0x15F8D5: void BM_parquet_read_io_compression<(cudf::io::io_type)0, (cudf::io::compression_type)2>(nvbench::state&, nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)0, cudf::io::io_type>, nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type> >) [clone .isra.0] (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/benchmarks/PARQUET_READER_NVBENCH)
==45341==    by 0x15FACE: void nvbench::tl::detail::foreach<nvbench::type_list<nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)0, cudf::io::io_type>, nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)0, cudf::io::io_type>, nvbench::enum_type<(cudf::io::compression_type)0, cudf::io::compression_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)1, cudf::io::io_type>, nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)1, cudf::io::io_type>, nvbench::enum_type<(cudf::io::compression_type)0, cudf::io::compression_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)2, cudf::io::io_type>, nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)2, cudf::io::io_type>, nvbench::enum_type<(cudf::io::compression_type)0, cudf::io::compression_type> > >, nvbench::runner<nvbench::benchmark<BM_parquet_read_io_compression_line_142, nvbench::type_list<nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)0, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)1, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)2, cudf::io::io_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type>, nvbench::enum_type<(cudf::io::compression_type)0, cudf::io::compression_type> > > > >::run_device(std::optional<nvbench::device_info> const&)::{lambda(auto:1)#1}, 0ul, 1ul, 2ul, 3ul, 4ul, 5ul>(std::integer_sequence<unsigned long, 0ul, 1ul, 2ul, 3ul, 4ul, 5ul>, nvbench::runner<nvbench::benchmark<BM_parquet_read_io_compression_line_142, nvbench::type_list<nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)0, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)1, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)2, cudf::io::io_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type>, nvbench::enum_type<(cudf::io::compression_type)0, cudf::io::compression_type> > > > >::run_device(std::optional<nvbench::device_info> const&)::{lambda(auto:1)#1}&&) [clone .isra.0] (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/benchmarks/PARQUET_READER_NVBENCH)
==45341==    by 0x161D09: nvbench::runner<nvbench::benchmark<BM_parquet_read_io_compression_line_142, nvbench::type_list<nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)0, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)1, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)2, cudf::io::io_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type>, nvbench::enum_type<(cudf::io::compression_type)0, cudf::io::compression_type> > > > >::run() (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/benchmarks/PARQUET_READER_NVBENCH)
==45341==    by 0x161DAC: nvbench::benchmark<BM_parquet_read_io_compression_line_142, nvbench::type_list<nvbench::type_list<nvbench::enum_type<(cudf::io::io_type)0, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)1, cudf::io::io_type>, nvbench::enum_type<(cudf::io::io_type)2, cudf::io::io_type> >, nvbench::type_list<nvbench::enum_type<(cudf::io::compression_type)2, cudf::io::compression_type>, nvbench::enum_type<(cudf::io::compression_type)0, cudf::io::compression_type> > > >::do_run() (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/benchmarks/PARQUET_READER_NVBENCH)
==45341==    by 0x12837E: main (in /home/wence/Documents/src/rapids/cudf/cpp/build/cuda-11.8.0/branch-23.06/release/benchmarks/PARQUET_READER_NVBENCH)
==45341== 
[ 45341][17:55:37:116091][warning] Running benchmarks without dropping the L3 cache; results may not reflect file IO throughput

[sdrp713]

I found that when setting the io type to Filepath, it is the source of the segfaults and cudaContext errors. I was wondering if the issue had some relation to kvikio, as this is the error I am seeing when running valgrind. I was also wondering if there is any way to disable using kvikio to see if that is the source of the issue?

==179002== 576 bytes in 1 blocks are possibly lost in loss record 35,124 of 35,671
==179002==    at 0x484DA83: calloc (in /usr/libexec/valgrind/vgpreload_memcheck-amd64-linux.so)
==179002==    by 0x40147D9: calloc (rtld-malloc.h:44)
==179002==    by 0x40147D9: allocate_dtv (dl-tls.c:375)
==179002==    by 0x40147D9: _dl_allocate_tls (dl-tls.c:634)
==179002==    by 0xD80B7B4: allocate_stack (allocatestack.c:430)
==179002==    by 0xD80B7B4: pthread_create@@GLIBC_2.34 (pthread_create.c:647)
==179002==    by 0xD57FDF3: __gthread_create (gthr-default.h:663)
==179002==    by 0xD57FDF3: std::thread::_M_start_thread(std::unique_ptr<std::thread::_State, std::default_delete<std::thread::_State> >, void (*)()) (thread.cc:172)
==179002==    by 0x5B455AB: kvikio::defaults::defaults() (in /home/raprabhu/miniconda3/envs/cudf_dev/lib/libcudf.so)
==179002==    by 0x5B4CAFE: cudf::io::file_sink::file_sink(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&) (in /home/raprabhu/miniconda3/envs/cudf_dev/lib/libcudf.so)
==179002==    by 0x5B4D593: cudf::io::data_sink::create(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&) (in /home/raprabhu/miniconda3/envs/cudf_dev/lib/libcudf.so)
==179002==    by 0x58D11DE: cudf::io::(anonymous namespace)::make_datasinks(cudf::io::sink_info const&) (in /home/raprabhu/miniconda3/envs/cudf_dev/lib/libcudf.so)
==179002==    by 0x58D1940: cudf::io::write_parquet(cudf::io::parquet_writer_options const&, rmm::cuda_stream_view) (in /home/raprabhu/miniconda3/envs/cudf_dev/lib/libcudf.so)
==179002==    by 0x1583BC: BM_parquet_read_io_compression(nvbench::state&) (in /home/raprabhu/cudf/cpp/build/benchmarks/PARQUET_READER_NVBENCH)
==179002==    by 0x165398: nvbench::runner<nvbench::benchmark<BM_parquet_read_io_compression_line_300, nvbench::type_list<> > >::run() (in /home/raprabhu/cudf/cpp/build/benchmarks/PARQUET_READER_NVBENCH)
==179002==    by 0x1655EC: nvbench::benchmark<BM_parquet_read_io_compression_line_300, nvbench::type_list<> >::do_run() (in /home/raprabhu/cudf/cpp/build/benchmarks/PARQUET_READER_NVBENCH)
==179002== 

==179002== 2,560 bytes in 4 blocks are possibly lost in loss record 35,368 of 35,671
==179002==    at 0x484DA83: calloc (in /usr/libexec/valgrind/vgpreload_memcheck-amd64-linux.so)
==179002==    by 0x40147D9: calloc (rtld-malloc.h:44)
==179002==    by 0x40147D9: allocate_dtv (dl-tls.c:375)
==179002==    by 0x40147D9: _dl_allocate_tls (dl-tls.c:634)
==179002==    by 0xD80B7B4: allocate_stack (allocatestack.c:430)
==179002==    by 0xD80B7B4: pthread_create@@GLIBC_2.34 (pthread_create.c:647)
==179002==    by 0x57D42149: ??? (in /home/raprabhu/miniconda3/envs/cudf_dev/targets/x86_64-linux/lib/libcufile.so.1.7.2)
==179002==    by 0x57D4521C: ??? (in /home/raprabhu/miniconda3/envs/cudf_dev/targets/x86_64-linux/lib/libcufile.so.1.7.2)
==179002==    by 0x57D31280: ??? (in /home/raprabhu/miniconda3/envs/cudf_dev/targets/x86_64-linux/lib/libcufile.so.1.7.2)
==179002==    by 0x57D54E4D: cuFileDriverOpen (in /home/raprabhu/miniconda3/envs/cudf_dev/targets/x86_64-linux/lib/libcufile.so.1.7.2)
==179002==    by 0x5B42C03: kvikio::cuFileAPI::cuFileAPI() (in /home/raprabhu/miniconda3/envs/cudf_dev/lib/libcudf.so)
==179002==    by 0x5B45766: kvikio::defaults::defaults() (in /home/raprabhu/miniconda3/envs/cudf_dev/lib/libcudf.so)
==179002==    by 0x5B4CAFE: cudf::io::file_sink::file_sink(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&) (in /home/raprabhu/miniconda3/envs/cudf_dev/lib/libcudf.so)
==179002==    by 0x5B4D593: cudf::io::data_sink::create(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&) (in /home/raprabhu/miniconda3/envs/cudf_dev/lib/libcudf.so)
==179002==    by 0x58D11DE: cudf::io::(anonymous namespace)::make_datasinks(cudf::io::sink_info const&) (in /home/raprabhu/miniconda3/envs/cudf_dev/lib/libcudf.so)

==179002== 12,288 bytes in 1 blocks are possibly lost in loss record 35,487 of 35,671
==179002==    at 0x4848899: malloc (in /usr/libexec/valgrind/vgpreload_memcheck-amd64-linux.so)
==179002==    by 0x4013DD9: malloc (rtld-malloc.h:56)
==179002==    by 0x4013DD9: allocate_dtv_entry (dl-tls.c:696)
==179002==    by 0x4013DD9: allocate_and_init (dl-tls.c:709)
==179002==    by 0x4013DD9: tls_get_addr_tail (dl-tls.c:907)
==179002==    by 0x401820B: __tls_get_addr (tls_get_addr.S:55)
==179002==    by 0x57D3595F: ??? (in /home/raprabhu/miniconda3/envs/cudf_dev/targets/x86_64-linux/lib/libcufile.so.1.7.2)
==179002==    by 0x57DF2C99: ??? (in /home/raprabhu/miniconda3/envs/cudf_dev/targets/x86_64-linux/lib/libcufile.so.1.7.2)
==179002==    by 0x57D9089A: ??? (in /home/raprabhu/miniconda3/envs/cudf_dev/targets/x86_64-linux/lib/libcufile.so.1.7.2)
==179002==    by 0x57D352A2: ??? (in /home/raprabhu/miniconda3/envs/cudf_dev/targets/x86_64-linux/lib/libcufile.so.1.7.2)
==179002==    by 0x57D55582: cuFileWrite (in /home/raprabhu/miniconda3/envs/cudf_dev/targets/x86_64-linux/lib/libcufile.so.1.7.2)
==179002==    by 0x5B4B1CD: kvikio::FileHandle::write(void const*, unsigned long, unsigned long, unsigned long) (in /home/raprabhu/miniconda3/envs/cudf_dev/lib/libcudf.so)
==179002==    by 0x5B4C184: std::_Function_handler<void (), BS::thread_pool::submit_task<kvikio::detail::submit_task<kvikio::FileHandle::pwrite(void const*, unsigned long, unsigned long, unsigned long, unsigned long)::{lambda(void const*, unsigned long, unsigned long, unsigned long)#3}, void*>(kvikio::FileHandle::pwrite(void const*, unsigned long, unsigned long, unsigned long, unsigned long)::{lambda(void const*, unsigned long, unsigned long, unsigned long)#3}, void*, unsigned long, unsigned long, unsigned long)::{lambda()#1}, unsigned long>(kvikio::detail::submit_task<kvikio::FileHandle::pwrite(void const*, unsigned long, unsigned long, unsigned long, unsigned long)::{lambda(void const*, unsigned long, unsigned long, unsigned long)#3}, void*>(kvikio::FileHandle::pwrite(void const*, unsigned long, unsigned long, unsigned long, unsigned long)::{lambda(void const*, unsigned long, unsigned long, unsigned long)#3}, void*, unsigned long, unsigned long, unsigned long)::{lambda()#1}&&)::{lambda()#1}>::_M_invoke(std::_Any_data const&) (in /home/raprabhu/miniconda3/envs/cudf_dev/lib/libcudf.so)
==179002==    by 0x5B38DC6: BS::thread_pool::worker(unsigned int, std::function<void ()> const&) (in /home/raprabhu/miniconda3/envs/cudf_dev/lib/libcudf.so)
==179002==    by 0xD57FE94: execute_native_thread_routine (thread.cc:104)

[GregoryKimball]

Thank you @sdrp713 for looking into this segfault and sharing some valgrind results. You might try setting LIBCUDF_CUFILE_POLICY=OFF (more info here). Also @madsbk would you please let us know if you have any ideas?

[madsbk]

Or try LIBCUDF_CUFILE_POLICY=KVIKIO KVIKIO_COMPAT_MODE=on, which make KvikIO skip the use of libcufile.so.

However, I think the valgrind outputs looks like innocent memory leaks at exit? Not something to trigger segfaults?

[sdrp713]

Ok, I tried setting LIBCUDF_CUFILE_POLICY=OFF and the seg fault was gone; however, when I set LIBCUDF_CUFILE_POLICY=KVIKIO KVIKIO_COMPAT_MODE=on, the seg fault came back

[madsbk]

I am not able to reproduce :/
I get some memory leaks but no access errors:

$ LIBCUDF_CUFILE_POLICY=KVIKIO KVIKIO_COMPAT_MODE=on valgrind cpp/build/benchmarks/PARQUET_READER_NVBENCH

... 

==235376== Warning: set address range perms: large range [0x106000000, 0x118000000) (noaccess)
==235376== 
==235376== HEAP SUMMARY:
==235376==     in use at exit: 2,271,605,821 bytes in 650,420 blocks
==235376==   total heap usage: 62,614,049 allocs, 61,963,629 frees, 77,697,641,434 bytes allocated
==235376== 
==235376== LEAK SUMMARY:
==235376==    definitely lost: 40 bytes in 1 blocks
==235376==    indirectly lost: 1,536 bytes in 33 blocks
==235376==      possibly lost: 446,282,686 bytes in 3,312 blocks
==235376==    still reachable: 1,825,321,559 bytes in 647,074 blocks
==235376==         suppressed: 0 bytes in 0 blocks
==235376== Rerun with --leak-check=full to see details of leaked memory
==235376== 
==235376== For lists of detected and suppressed errors, rerun with: -s
==235376== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 0 from 0)

Can you try running until it segfaults with valgrind?