stat: use long doubles to identify latency percentiles
authorVincent Fu <vincent.fu@wdc.com>
Tue, 19 Feb 2019 21:44:07 +0000 (16:44 -0500)
committerJens Axboe <axboe@kernel.dk>
Thu, 21 Feb 2019 17:55:30 +0000 (10:55 -0700)
In some cases, the 100th percentile latency is not correctly identified
because of problems with double precision floating point arithmetic.
Use long doubles instead in the while loop condition to reduce the
likelihood of encountering this problem.

Also, print an error message when latency percentiles are not
successfully identified.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
stat.c

diff --git a/stat.c b/stat.c
index c1f46e1..66a13bc 100644 (file)
--- a/stat.c
+++ b/stat.c
@@ -170,7 +170,7 @@ unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
        is_last = false;
        for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
                sum += io_u_plat[i];
-               while (sum >= (plist[j].u.f / 100.0 * nr)) {
+               while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
                        assert(plist[j].u.f <= 100.0);
 
                        ovals[j] = plat_idx_to_val(i);
@@ -187,6 +187,9 @@ unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
                }
        }
 
+       if (!is_last)
+               log_err("fio: error calculating latency percentiles\n");
+
        *output = ovals;
        return len;
 }