From: jrchyang <yujrchyang@outlook.com>
Date: Tue, 12 Sep 2023 09:31:28 +0000 (+0000)
Subject: osd: fix: slow scheduling when item_cost is large
X-Git-Tag: v17.2.7~37^2
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=refs%2Fpull%2F53860%2Fhead;p=ceph.git

osd: fix: slow scheduling when item_cost is large

We use the iops and bandwidth tested by
`ceph tell osd.0 bench 10737418240 204800 204800 100`
to verify the QoS function. iops was 400 and bandwidth was 80MiB/s.
When osd_mclock_scheduler_client_lim is set to 1,
the sequential write bandwidth is only half of the capacity.
Therefore, we believe that it should not unconditionally increase
osd_bandwidth_cost_per_io for each IO, but take the maximum of the two.

Fixes: https://tracker.ceph.com/issues/62812
co-author: yanghonggang <yanghonggang_yewu@cmss.chinamobile.com>
co-author: zhangjianwei <zhangjianwei2_yewu@cmss.chinamobile.com>
Signed-off-by: Jrchyang Yu <yuzhiqiang_yewu@cmss.chinamobile.com>
(cherry picked from commit 65ebf2d6f1b41c93ee2e170437780846d20f5480)
---

diff --git a/src/osd/scheduler/mClockScheduler.cc b/src/osd/scheduler/mClockScheduler.cc
index 4b30766796c6..876fed89e2b7 100644
--- a/src/osd/scheduler/mClockScheduler.cc
+++ b/src/osd/scheduler/mClockScheduler.cc
@@ -338,8 +338,7 @@ uint32_t mClockScheduler::calc_scaled_cost(int item_cost)
       item_cost));
   auto cost_per_io = static_cast<uint32_t>(osd_bandwidth_cost_per_io);
 
-  // Calculate total scaled cost in bytes
-  return cost_per_io + cost;
+  return std::max<uint32_t>(cost, cost_per_io);
 }
 
 void mClockScheduler::update_configuration()