LU-6370 osc: disable to control per-OSC LRU budget

It will produce a problem where a system has lots of OSTs but
only a few of them are involved in doing read on the client side.
In that case, the per-OSC budget will cause the read ahead pages to
be dropped aggressively, which results in significantly performance
degradation.

This patch disabled per-OSC LRU budget. However, this is not a real
fix. A real fix should feed back memory pressure from LRU on the OSC
layer to the LLITE layer, so that read ahead algorithm can take that
information into consideration when it decides read ahead window size.

Signed-off-by: Jinshan Xiong <jinshan.xiong@intel.com>
Change-Id: Ifa81d8515a345b389d74af971582b2a3b1761546
Reviewed-on: http://review.whamcloud.com/14347
Tested-by: Jenkins
Reviewed-by: Patrick Farrell <paf@cray.com>
Tested-by: Maloo <hpdd-maloo@intel.com>
Reviewed-by: Bobi Jam <bobijam@hotmail.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>

diff --git a/lustre/osc/osc_page.c b/lustre/osc/osc_page.c
index f17c4e9..dd83ef9 100644 (file)
--- a/lustre/osc/osc_page.c
+++ b/lustre/osc/osc_page.c
@@ -510,8 +510,11 @@ static int osc_cache_too_much(struct client_obd *cli)
                        return lru_shrink_max;
                else if (pages >= budget / 2)
                        return lru_shrink_min;
+#if 0
        } else if (pages >= budget * 2)
                return lru_shrink_min;
+#endif
+       }
        return 0;
 }