On 12/03/2015 10:19 AM, Avi Kivity wrote:
On 12/03/2015 07:01 PM, Jens Axboe wrote:
On 12/01/2015 04:04 AM, Avi Kivity wrote:
Sometimes you want to run a set of experiments on a disk, varying a
parameter between tests (in my case, iodepth, but buffer size is also a
good candidate). You then want to present the results in a nice graph.
I wrote a small wrapper around fio to do this
(https://github.com/avikivity/diskplorer), but it occurs to me that
generalized support for both in fio would be much more useful.
Possibly, you'd define a job as a template:
[aio-read]
template_start=1
template_end=100
template_step=1
(or template_ratio=1.05 for exponential growth)
iodepth=template_variable
(it's just possible that someone can come up with better syntax).
A few more options in the global section can then cause a graph to be
generated.
It'd be great to integrate this into fio, as graphing results is
something that most people want to do. Any chance you would be willing
to try and hash that out?
I'd love to say yes, but no.
:-)
btw, a fast disk can easily saturate a single core using libaio, so a
multithreaded libaio ioengine would be welcome (I am currently emulating
it using multiple jobs and new_group).
In the context of fio, that doesn't make a lot of sense. A job in fio
is, by definition, either a thread or a process that does IO. So if
you want more threads banging on a device, then you'd add more jobs.
If multiple threads shared on aio context, then we'd also potentially
see contention on that part. If you just use more jobs, then each gets
an aio context as well.
If your jobs are generated via a template, as above, then this is hard
to do. For iodepth=1 you want one job with iodepth=1. For iodepth=64
you want 8 jobs (one per core) each with an iodepth=8; otherwise a fast
SSD will overwhelm a single core.
Perhaps the job specification can be modified so that it auto-generates
subjobs. In the specification, there is one entry, but fio sees 8 (or
1, when the template sets iodepth=1), and reports them via a group.
[aio-read]
template_start=1
template_end=100
template_step=1
(or template_ratio=1.05 for exponential growth)
subjobs=(min(template_variable, core_count))
iodepth=(template_variable / subjobs)
(the above doesn't cope will with an iodepth that doesn't divide into
your core_count; displorer will generate subjobs with different iodepth
for this)
For purely automated or templated, yeah, it's not ideal. But some of
this is highly setup specific. For QD=64, 2 threads at QD=32 might be
the best option. Or 8/8, perhaps. Sometimes it's a tradeoff between
throwing CPU cycles at it to squeeze out the last drop of performance,
sometimes (eg on nvme), you need "just enough" threads to reach max
performance, since things are mostly 100% parallelized on both the
submission and completion path.
Fio does support thread offload (io_submit_mode=offload), which was
added not for performance reasons, but because it's important to capture
the true latency of the device in case of device backup. So the
framework is in place to do that - which was the harder part. You might
want to take a look at that. It would need slight modifications for your
use case, but it'd be a good general addition.
--
Jens Axboe
--
To unsubscribe from this list: send the line "unsubscribe fio" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
