RE: last_block() vs. min_bs when > blockalign

* RE: last_block() vs. min_bs when > blockalign
@ 2015-01-27 22:42 Justin Eno (jeno)
  2015-01-28 16:11 ` Jens Axboe
  0 siblings, 1 reply; 4+ messages in thread
From: Justin Eno (jeno) @ 2015-01-27 22:42 UTC (permalink / raw)
  To: fio

[-- Attachment #1: Type: text/plain, Size: 4030 bytes --]

The attached updated patch produces the desired behavior.  If min_bs is larger than blockalign, last_block() now returns a suitable value.

Thanks,
Justin

-----Original Message-----
From: Justin Eno (jeno) 
Sent: Tuesday, January 27, 2015 1:35 PM
To: 'fio@vger.kernel.org'
Subject: RE: last_block() vs. min_bs when > blockalign

Please disregard the patch - it incorrectly limits the target write range.  I will work on a proper fix.  Sorry for the noise.

-Justin

-----Original Message-----
From: Justin Eno (jeno) 
Sent: Tuesday, January 27, 2015 12:00 PM
To: 'fio@vger.kernel.org'
Subject: last_block() vs. min_bs when > blockalign

Hi Jens and all,

I have found that when performing mixed-sized random writes, fio ceases writing early due to last_block() not respecting min_bs.
The attached workload demonstrates the issue.  When run with io debugging enabled, "failed getting buflen" is reported, and the written data falls short of the total requested (in this case, 10M), e.g., ###############################################
[jeno@rhel fio-clean]$ ./fio --debug=io examples/random-fill.fio
fio: Any use of blockalign= turns off randommap
fio: set debug option io
io       25214 load ioengine sync
write-phase: (g=0): rw=randwrite, bs=1K-5K/1K-5K/1K-5K, ioengine=sync, iodepth=1 fio-2.2.5-3-g209e Starting 1 thread
write-phase: Laying out IO file(s) (1 file(s) / 5MB)
io       25214 invalidate cache datafile.tmp: 0/5242880
io       25214 invalidate cache datafile.tmp: 0/5242880
io       25214 fill_io_u: io_u 0x7fe49c003d40: off=315904/len=1024/ddir=1/datafile.tmp
io       25214 prep: io_u 0x7fe49c003d40: off=315904/len=1024/ddir=1/datafile.tmp
io       25214 ->prep(0x7fe49c003d40)=0
io       25214 queue: io_u 0x7fe49c003d40: off=315904/len=1024/ddir=1/datafile.tmp
io       25214 io complete: io_u 0x7fe49c003d40: off=315904/len=1024/ddir=1/datafile.tmp
...
io       25214 fill_io_u: io_u 0x7fe49c003d40: off=4930560/len=5120/ddir=1/datafile.tmp
io       25214 prep: io_u 0x7fe49c003d40: off=4930560/len=5120/ddir=1/datafile.tmp
io       25214 ->prep(0x7fe49c003d40)=0
io       25214 queue: io_u 0x7fe49c003d40: off=4930560/len=5120/ddir=1/datafile.tmp
io       25214 io complete: io_u 0x7fe49c003d40: off=4930560/len=5120/ddir=1/datafile.tmp
io       25214 io_u 0x7fe49c003d40, failed getting buflen
io       25214 io_u 0x7fe49c003d40, setting file failed
io       25214 get_io_u failed
io       25214 close ioengine sync
io       25214 free ioengine sync

write-phase: (groupid=0, jobs=1): err= 0: pid=25281: Tue Jan 27 10:33:50 2015
  write: io=4335.0KB, bw=585007B/s, iops=195, runt=  7588msec
    clat (usec): min=762, max=17788, avg=5100.38, stdev=2509.05
     lat (usec): min=764, max=17790, avg=5101.80, stdev=2509.06
    clat percentiles (usec):
     |  1.00th=[  868],  5.00th=[ 1240], 10.00th=[ 1704], 20.00th=[ 2544],
     | 30.00th=[ 3440], 40.00th=[ 4192], 50.00th=[ 5024], 60.00th=[ 5984],
     | 70.00th=[ 6752], 80.00th=[ 7648], 90.00th=[ 8384], 95.00th=[ 8896],
     | 99.00th=[ 9280], 99.50th=[10304], 99.90th=[17792], 99.95th=[17792],
     | 99.99th=[17792]
    bw (KB  /s): min=  520, max=  640, per=99.98%, avg=570.87, stdev=35.30
    lat (usec) : 1000=2.76%
    lat (msec) : 2=11.25%, 4=22.76%, 10=62.63%, 20=0.61%
  cpu          : usr=0.12%, sys=0.28%, ctx=1486, majf=0, minf=1
  IO depths    : 1=100.0%, 2=0.0%, 4=0.0%, 8=0.0%, 16=0.0%, 32=0.0%, >=64=0.0%
     submit    : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
     complete  : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
     issued    : total=r=0/w=1485/d=0, short=r=0/w=0/d=0, drop=r=0/w=0/d=0
     latency   : target=0, window=0, percentile=100.00%, depth=1

Run status group 0 (all jobs):
  WRITE: io=4335KB, aggrb=571KB/s, minb=571KB/s, maxb=571KB/s, mint=7588msec, maxt=7588msec ...
###############################################

The attached patch changes last_block() to respect min_bs.

Thanks,
Justin


[-- Attachment #2: last_block_v2.patch --]
[-- Type: application/octet-stream, Size: 957 bytes --]

From f4f9dd6eb3b81f2cb79aa3c41fd4d131cca91b6c Mon Sep 17 00:00:00 2001
From: Justin Eno <jeno@micron.com>
Date: Tue, 27 Jan 2015 14:23:20 -0800
Subject: [PATCH] Better accommodate random writes larger than blockalign

fill_io_u() fails prematurely if the randomly-chosen offset satisfies
blockalign but not min_bs, i.e., the offset lies too near the end of
the target region.  This change honors both parameters.

Signed-off-by: Justin Eno <jeno@micron.com>
---
 io_u.c |    3 +++
 1 files changed, 3 insertions(+), 0 deletions(-)

diff --git a/io_u.c b/io_u.c
index 5971d78..77a3abf 100644
--- a/io_u.c
+++ b/io_u.c
@@ -68,6 +68,9 @@ static uint64_t last_block(struct thread_data *td, struct fio_file *f,
 	if (td->o.zone_range)
 		max_size = td->o.zone_range;
 
+	if (td->o.min_bs[ddir] > td->o.ba[ddir])
+		max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
+
 	max_blocks = max_size / (uint64_t) td->o.ba[ddir];
 	if (!max_blocks)
 		return 0;
-- 
1.7.1


^ permalink raw reply related	[flat|nested] 4+ messages in thread