toddlipcon

Before and after comparing the CPU usage of the kudu tserver,
scanning the ~3.4B rows of the l_shipmode column from TPCH lineitem:

Before
=============
 Performance counter stats for 'numactl -m 0 -N 0 build/release/bin/kudu-tserver -fs-wal-dir /data/1/todd/tpch-kudu-data/ts-0/wal/ -fs-data-dirs /data/1/todd/tpch-kudu-data/ts-0/data/':

        137,685.93 msec task-clock                #    3.356 CPUs utilized          
           106,215      context-switches          #    0.771 K/sec                  
               706      cpu-migrations            #    0.005 K/sec                  

toddlipcon

Before and after comparing the CPU usage of the kudu tserver,
scanning the ~3.4B rows of the l_shipmode column from TPCH lineitem:

Before
=============
 Performance counter stats for 'numactl -m 0 -N 0 build/release/bin/kudu-tserver -fs-wal-dir /data/1/todd/tpch-kudu-data/ts-0/wal/ -fs-data-dirs /data/1/todd/tpch-kudu-data/ts-0/data/':

        137,685.93 msec task-clock                #    3.356 CPUs utilized          
           106,215      context-switches          #    0.771 K/sec                  
               706      cpu-migrations            #    0.005 K/sec                  

toddlipcon

Legend
===================
Methods:
  fd:      client sends a file descriptor for the server to fill
  s-write: server writes data the passed fd using write() syscall
  s-mmap:  server mmaps() the passed fd and writes using memset()
  s-reuse: server holds on to the mmapped region across calls
  c-reuse: client holds on to the mmapped region across calls
  sidecar: server sends back data over unix socket
  nt:      server uses nontemporal store instructions to write

toddlipcon

Legend
===================
Methods:
  fd:      client sends a file descriptor for the server to fill
  s-write: server writes data the passed fd using write() syscall
  s-mmap:  server mmaps() the passed fd and writes using memset()
  s-reuse: server holds on to the mmapped region across calls
  c-reuse: client holds on to the mmapped region across calls
  sidecar: server sends back data over unix socket
  nt:      server uses nontemporal store instructions to write

toddlipcon

Legend
===================
Methods:
  fd:      client sends a file descriptor for the server to fill
  s-write: server writes data the passed fd using write() syscall
  s-mmap:  server mmaps() the passed fd and writes using memset()
  s-reuse: server holds on to the mmapped region across calls
  c-reuse: client holds on to the mmapped region across calls
  sidecar: server sends back data over unix socket
  nt:      server uses nontemporal store instructions to write

toddlipcon

todd@va1022:/tmp$ cat test.cc
int main() {
  return __builtin_cpu_supports("bmi") ? 0 : 1;
}
todd@va1022:/tmp$ g++ ./test.cc -O2   ; ./a.out  && echo "supports bmi" || echo "no bmi"
no bmi
todd@va1022:/tmp$ g++ ./test.cc -O2 /opt/rh/devtoolset-9/root/usr/lib/gcc/x86_64-redhat-linux/9/libgcc.a  ; ./a.out  && echo "supports bmi" || echo "no bmi"
supports bmi

toddlipcon

Legend
===================
Methods:
  fd:      client sends a file descriptor for the server to fill
  s-write: server writes data the passed fd using write() syscall
  s-mmap:  server mmaps() the passed fd and writes using memset()
  s-reuse: server holds on to the mmapped region across calls
  c-reuse: client holds on to the mmapped region across calls
  sidecar: server sends back data over unix socket
  nt:      server uses nontemporal store instructions to write

toddlipcon

Legend
===================
Methods:
  fd:      client sends a file descriptor for the server to fill
  s-write: server writes data the passed fd using write() syscall
  s-mmap:  server mmaps() the passed fd and writes using memset()
  s-reuse: server holds on to the mmapped region across calls
  c-reuse: client holds on to the mmapped region across calls
  sidecar: server sends back data over unix socket
  nt:      server uses nontemporal store instructions to write

toddlipcon

Legend
===================
Methods:
  fd:      client sends a file descriptor for the server to fill
  s-write: server writes data the passed fd using write() syscall
  s-mmap:  server mmaps() the passed fd and writes using memset()
  s-reuse: server holds on to the mmapped region across calls
  c-reuse: client holds on to the mmapped region across calls
  sidecar: server sends back data over unix socket
  nt:      server uses nontemporal store instructions to write

toddlipcon

// Curiously-recurring-template-pattern base class for private arrow data.
//
// This is similar to RefCountedThreadSafe<T> except that it participates in the
// C-style release callback expected by the Arrow C interface.
template<class T>
struct ArrowPrivateDataBase {
  ~ArrowPrivateDataBase() {
    DCHECK_EQ(0, ref_count);
  }