Instead of locking around every request, use explicit and (slightly) more fine-grained, and change the thread model to parallel. The same change is made to the memory plugin and the data plugin. This improves performance slightly. Using fio with 8 threads and multi-conn enabled with -C 8: Before: read: IOPS=103k, BW=401MiB/s (420MB/s)(46.0GiB/120002msec) write: IOPS=103k, BW=401MiB/s (420MB/s)(46.0GiB/120002msec) After: read: IOPS=112k, BW=437MiB/s (458MB/s)(51.2GiB/120001msec) write: IOPS=112k, BW=437MiB/s (458MB/s)(51.2GiB/120001msec) For comparison: The memory plugin implemented using a simple malloc instead of a sparse array: read: IOPS=133k, BW=518MiB/s (544MB/s)(60.7GiB/120002msec) write: IOPS=133k, BW=518MiB/s (543MB/s)(60.7GiB/120002msec) Directly running fio against /dev/shm: read: IOPS=1018k, BW=3978MiB/s (4171MB/s)(466GiB/120001msec) write: IOPS=1018k, BW=3979MiB/s (4172MB/s)(466GiB/120001msec) --- plugins/data/data.c | 22 +++++++++++++++++++--- plugins/memory/memory.c | 21 +++++++++++++++++++-- 2 files changed, 38 insertions(+), 5 deletions(-) diff --git a/plugins/data/data.c b/plugins/data/data.c index 11a7b69..1d4c5b9 100644 --- a/plugins/data/data.c +++ b/plugins/data/data.c @@ -39,6 +39,8 @@ #include <inttypes.h> #include <string.h> +#include <pthread.h> + #if defined(HAVE_GNUTLS) && defined(HAVE_GNUTLS_BASE64_DECODE2) #include <gnutls/gnutls.h> #endif @@ -56,8 +58,11 @@ static int64_t size = -1; /* Size of data specified on the command line. */ static int64_t data_size = -1; -/* Sparse array. */ +/* Sparse array - the lock must be held when accessing this from + * connected callbacks. + */ static struct sparse_array *sa; +static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER; /* Debug directory operations (-D data.dir=1). */ int data_debug_dir; @@ -305,7 +310,7 @@ data_dump_plugin (void) #endif } -#define THREAD_MODEL NBDKIT_THREAD_MODEL_SERIALIZE_ALL_REQUESTS +#define THREAD_MODEL NBDKIT_THREAD_MODEL_PARALLEL /* No meaning, just used as the address for the handle. */ static int dh; @@ -335,7 +340,9 @@ data_can_multi_conn (void *handle) static int data_pread (void *handle, void *buf, uint32_t count, uint64_t offset) { + pthread_mutex_lock (&lock); sparse_array_read (sa, buf, count, offset); + pthread_mutex_unlock (&lock); return 0; } @@ -343,14 +350,21 @@ data_pread (void *handle, void *buf, uint32_t count, uint64_t offset) static int data_pwrite (void *handle, const void *buf, uint32_t count, uint64_t offset) { - return sparse_array_write (sa, buf, count, offset); + int r; + + pthread_mutex_lock (&lock); + r = sparse_array_write (sa, buf, count, offset); + pthread_mutex_unlock (&lock); + return r; } /* Zero. */ static int data_zero (void *handle, uint32_t count, uint64_t offset, int may_trim) { + pthread_mutex_lock (&lock); sparse_array_zero (sa, count, offset); + pthread_mutex_unlock (&lock); return 0; } @@ -358,7 +372,9 @@ data_zero (void *handle, uint32_t count, uint64_t offset, int may_trim) static int data_trim (void *handle, uint32_t count, uint64_t offset) { + pthread_mutex_lock (&lock); sparse_array_zero (sa, count, offset); + pthread_mutex_unlock (&lock); return 0; } diff --git a/plugins/memory/memory.c b/plugins/memory/memory.c index 0084401..e27e127 100644 --- a/plugins/memory/memory.c +++ b/plugins/memory/memory.c @@ -42,6 +42,8 @@ #include <errno.h> #include <assert.h> +#include <pthread.h> + #include <nbdkit-plugin.h> #include "sparse.h" @@ -52,7 +54,11 @@ static int64_t size = 0; /* Debug directory operations (-D memory.dir=1). */ int memory_debug_dir; +/* Sparse array - the lock must be held when accessing this from + * connected callbacks. + */ static struct sparse_array *sa; +static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER; static void memory_load (void) @@ -106,7 +112,7 @@ memory_open (int readonly) return NBDKIT_HANDLE_NOT_NEEDED; } -#define THREAD_MODEL NBDKIT_THREAD_MODEL_SERIALIZE_ALL_REQUESTS +#define THREAD_MODEL NBDKIT_THREAD_MODEL_PARALLEL /* Get the disk size. */ static int64_t @@ -126,7 +132,9 @@ memory_can_multi_conn (void *handle) static int memory_pread (void *handle, void *buf, uint32_t count, uint64_t offset) { + pthread_mutex_lock (&lock); sparse_array_read (sa, buf, count, offset); + pthread_mutex_unlock (&lock); return 0; } @@ -134,14 +142,21 @@ memory_pread (void *handle, void *buf, uint32_t count, uint64_t offset) static int memory_pwrite (void *handle, const void *buf, uint32_t count, uint64_t offset) { - return sparse_array_write (sa, buf, count, offset); + int r; + + pthread_mutex_lock (&lock); + r = sparse_array_write (sa, buf, count, offset); + pthread_mutex_unlock (&lock); + return r; } /* Zero. */ static int memory_zero (void *handle, uint32_t count, uint64_t offset, int may_trim) { + pthread_mutex_lock (&lock); sparse_array_zero (sa, count, offset); + pthread_mutex_unlock (&lock); return 0; } @@ -149,7 +164,9 @@ memory_zero (void *handle, uint32_t count, uint64_t offset, int may_trim) static int memory_trim (void *handle, uint32_t count, uint64_t offset) { + pthread_mutex_lock (&lock); sparse_array_zero (sa, count, offset); + pthread_mutex_unlock (&lock); return 0; } -- 2.19.2