GUASI update

The GUASI library has been changed to avoid using pthread defaults stacks
(that burns 8MB of VM area for each thread). Also, a new paramter has been
added to guasi_create(), to allow the caller to specify a stack size.
I noticed a different behaviour of FIO in randrw mode, from GUASI to libaio.
In GUASI, the amount of data read is about the same +-2-3% of the amount
of data written. Whereas in libaio, there can be a huge difference between
the two (up to 2-3 times the other). This is consistent.
Besides that, GUASI performance follows pretty much the libaio one, up to
an io-depth of 1000-1500, after that shows about 10-15% less performance.

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>

diff --git a/engines/guasi.c b/engines/guasi.c
index 34b42e71454916ffa0492d378bb60bbac3e400e4..925c364a83019b84ac01d2136bc9a767e68cdf55 100644 (file)
--- a/engines/guasi.c
+++ b/engines/guasi.c
@@ -26,7 +26,7 @@
 
 #define GFIO_MIN_THREADS 32
 #ifndef GFIO_MAX_THREADS
-#define GFIO_MAX_THREADS 280
+#define GFIO_MAX_THREADS 2000
 #endif
 
 #include <guasi.h>
@@ -108,9 +108,12 @@ static int fio_guasi_getevents(struct thread_data *td, int min, int max,
                guasi_req_free(ld->reqs[n]);
        n = 0;
        do {
-               r = guasi_fetch(ld->hctx, ld->reqs + n, max - n, timeo);
-               if (r < 0)
+               r = guasi_fetch(ld->hctx, ld->reqs + n, min - n,
+                               max - n, timeo);
+               if (r < 0) {
+                       fprintf(stderr, "guasi_fetch() FAILED! (%d)\n", r);
                        break;
+               }
                n += r;
                if (n >= min)
                        break;
@@ -134,15 +137,18 @@ static int fio_guasi_queue(struct thread_data *td, struct io_u *io_u)
        return FIO_Q_QUEUED;
 }
 
-static void fio_guasi_queued(struct thread_data *td, struct io_u **io_us,
-                            unsigned int nr)
+static void fio_guasi_queued(struct thread_data *td, struct io_u **io_us, int nr)
 {
+       int i;
+       struct io_u *io_u;
        struct timeval now;
-       struct io_u *io_u = io_us[nr];
 
        fio_gettime(&now, NULL);
-       memcpy(&io_u->issue_time, &now, sizeof(now));
-       io_u_queued(td, io_u);
+       for (i = 0; i < nr; i++) {
+               io_u = io_us[i];
+               memcpy(&io_u->issue_time, &now, sizeof(now));
+               io_u_queued(td, io_u);
+       }
 }
 
 static int fio_guasi_commit(struct thread_data *td)
@@ -172,14 +178,13 @@ static int fio_guasi_commit(struct thread_data *td)
                        fprintf(stderr, "fio_guasi_commit() FAILED: unknow request %d\n",
                                io_u->ddir);
                }
-               if (io_u->greq != NULL)
-                       fio_guasi_queued(td, ld->io_us, i);
-               else {
+               if (io_u->greq == NULL) {
                        fprintf(stderr, "fio_guasi_commit() FAILED: submit failed (%s)\n",
                                strerror(errno));
                        return -1;
                }
        }
+       fio_guasi_queued(td, ld->io_us, i);
        ld->queued_nr = 0;
        GDBG_PRINT(("fio_guasi_commit() -> %d\n", i));
 
@@ -223,7 +228,7 @@ static int fio_guasi_init(struct thread_data *td)
        maxthr = td->o.iodepth > GFIO_MIN_THREADS ? td->o.iodepth: GFIO_MIN_THREADS;
        if (maxthr > GFIO_MAX_THREADS)
                maxthr = GFIO_MAX_THREADS;
-       if ((ld->hctx = guasi_create(GFIO_MIN_THREADS, maxthr, 1)) == NULL) {
+       if ((ld->hctx = guasi_create(GFIO_MIN_THREADS, maxthr, 1, 0)) == NULL) {
                td_verror(td, errno, "guasi_create");
                free(ld);
                return 1;