On 02/04/2014 01:08 PM, Gruher, Joseph R wrote:
-----Original Message-----
From: Gregory Farnum [mailto:greg@xxxxxxxxxxx]
Sent: Tuesday, February 04, 2014 9:46 AM
To: Gruher, Joseph R
Cc: Mark Nelson; ceph-users@xxxxxxxxxxxxxx; Ilya Dryomov
Subject: Re: Low RBD Performance
On Tue, Feb 4, 2014 at 9:29 AM, Gruher, Joseph R
<joseph.r.gruher@xxxxxxxxx> wrote:
-----Original Message-----
From: ceph-users-bounces@xxxxxxxxxxxxxx [mailto:ceph-users-
bounces@xxxxxxxxxxxxxx] On Behalf Of Mark Nelson
Sent: Monday, February 03, 2014 6:48 PM
To: ceph-users@xxxxxxxxxxxxxx
Subject: Re: Low RBD Performance
On 02/03/2014 07:29 PM, Gruher, Joseph R wrote:
Hi folks-
I'm having trouble demonstrating reasonable performance of RBDs.
I'm running Ceph 0.72.2 on Ubuntu 13.04 with the 3.12 kernel. I
have four dual-Xeon servers, each with 24GB RAM, and an Intel 320
SSD for journals and four WD 10K RPM SAS drives for OSDs, all
connected with an LSI 1078. This is just a lab experiment using
scrounged hardware so everything isn't sized to be a Ceph cluster,
it's just what I have lying around, but I should have more than
enough CPU and memory
resources.
Everything is connected with a single 10GbE.
When testing with RBDs from four clients (also running Ubuntu 13.04
with
3.12 kernel) I am having trouble breaking 300 IOPS on a 4KB random
read or write workload (cephx set to none, replication set to one).
IO is generated using FIO from four clients, each hosting a single
1TB RBD, and I've experimented with queue depths and increasing the
number of RBDs without any benefit. 300 IOPS for a pool of 16 10K
RPM HDDs seems quite low, not to mention the journal should provide
a good boost on write workloads. When I run a 4KB object write
workload in Cosbench I can approach 3500 Obj/Sec which seems more
reasonable.
Sample FIO configuration:
[global]
ioengine=libaio
direct=1
ramp_time=300
runtime=300
[4k-rw]
description=4k-rw
filename=/dev/rbd1
rw=randwrite
bs=4k
stonewall
I use --iodepth=X on the FIO command line to set the queue depth
when testing.
I notice in the FIO output despite the iodepth setting it seems to
be reporting an IO depth of only 1, which would certainly help
explain poor performance, but I'm at a loss as to why, I wonder if
it could be something specific to RBD behavior, like I need to use a
different IO engine to establish queue depth.
IO depths : 1=200.0%, 2=0.0%, 4=0.0%, 8=0.0%, 16=0.0%, 32=0.0%,
=64=0.0%
Any thoughts appreciated!
Interesting results with the io depth at 1. I Haven't seen that
behaviour when using libaio, direct=1, and higher io depths. Is this kernel
RBD or QEMU/KVM?
If it's QEMU/KVM, is it the libvirt driver?
Certainly 300 IOPS is low for that kind of setup compared to what
we've seen for RBD on other systems (especially with 1x replication).
Given that you are seeing more reasonable performance with RGW, I
guess I'd look at a couple
things:
- Figure out why fio is reporting queue depth = 1
Yup, I agree, I will work on this and report back. First thought is to try
specifying the queue depth in the FIO workload file instead of on the
command line.
- Does increasing the num jobs help (ie get concurrency another way)?
I will give this a shot.
- Do you have enough PGs in the RBD pool?
I should, for 16 OSDs and no replication I use 2048 PGs/PGPs (100 * 16 / 1
rounded up to power of 2).
- Are you using the virtio driver if QEMU/KVM?
No virtualization, clients are bare metal using kernel RBD.
I believe that directIO via the kernel client will go all the way to the OSDs and
to disk before returning. I imagine that something in the stack is preventing
the dispatch from actually happening asynchronously in that case, and the
reason you're getting 300 IOPS is because your total RTT is about 3 ms with
that code...
Ilya, is that assumption of mine correct? One thing that occurs to me is that for
direct IO it's fair to use the ack instead of on-disk response from the OSDs,
although that would only help us for people using btrfs.
-Greg
Software Engineer #42 @ http://inktank.com | http://ceph.com
Ultimately this seems to be an FIO issue. If I use "--iodepth X" or "--iodepth=X" on the FIO command line I always get queue depth 1. After switching to specifying "iodepth=X" in the body of the FIO workload file I do get the desired queue depth and I can immediately see performance is much higher (a full re-test is underway, I can share some results when complete if anyone is curious). This seems to have effectively worked around the problem, although I'm still curious why the command line parameters don't have the desired effect. Thanks for the responses!
Strange! I do most of our testing using the command line parameters as
well. What version of fio are you using? Maybe there is a bug. For
what it's worth, I'm using --iodepth=X, and fio version 1.59 from the
Ubuntu precise repository.
Mark
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
