Upon further reflection, there is no need to be hasty. My concern is simply interminable discussion. Bounded discussion OTOH sounds great. How about:

1. Open discussion from now until Jan 31 to define goals, identify concerns, etc. 
2. In Feb, people can submit benchmark proposals.  Two months to discuss and refine these. 
3. If consensus not reached by April 1, vote. 
4. Survey the top 100 sites using the chosen benchmark proposal




On Nov 23, 2016, at 12:37 AM, John Bent <John.Bent@seagategov.com> wrote:

PLFS will do horribly in this test because it will completely fail the IOR hard and it will do horribly on both mdtest hard and easy.  If someone fixes it (like maybe the Mohror burst-fs stuff), then that seems like fair game.

I think we’re at a bit of an impasse in terms of whether to use IOR and mdtest or try to create new and better benchmarks.  I fear that an attempt to create new and better benchmarks is doomed to failure.  We can never make everyone happy.  Dongarra went with Linpack.  It’s universally reviled yet Top 500 is massively successful.

Now, here’s a question: do we want to be successful or do we want to advance science?  Because maybe they are different things.  Has Top 500 with an imperfect benchmark advanced science?

In terms of our impasse, how about we spend six weeks trying to define a more well-liked benchmark?  If we fail, we go ahead with IOR and mdtest.  How do we determine failure/success?  I guarantee we won’t find consensus.  Maybe six weeks to discuss and then we submit candidates and vote?

I also like my idea of surveying the Top 100.  I suspect if we propose IOR hard, IOR easy, mdtest hard, mdtest easy to them today, we will get 30 of them that say they will do it.  I suspect if we spend six months discussing and then survey them again with whatever we agree upon, we will get 30 of them that say they will do it.  And I think 30 is enough that we should do it because 30 is large enough that the rest will follow.  Heck, I think ORNL alone is enough that the rest will follow.

But if our goal is not just to succeed, but actually to advance science then I’m willing to spend six weeks and see if we converge on something better than IOR and mdtest.  My prediction (which I think Andreas shares)?  We won’t.

Sarp and others who haven’t yet spoken, can you please weigh in on whether:

A. We should survey immediately with IOR and mdtest
B. We should spend six weeks trying to find something better
C. Some other path?

By the way, I liked Dean’s suggestion that we change the benchmark every year in terms of how well it would advance science.  But it absolutely terrifies me as a practical matter.  This discussion every year?  Plus I think sites are less likely to participate if they have to learn how to run, and tune for, a different benchmark every year.

Also, replies to Jay inline:

On Nov 22, 2016, at 5:03 PM, Lofstead, Gerald F II <gflofst@sandia.gov> wrote:

Sarp: Can you share any of the materials from your previous effort?
I have a few other comments to add in:
1. A better metadata testing tool is a great idea. Let’s focus on forward looking tools rather than clinging to old tools. My concern is how well can we avoid “gaming” the new tool. mdtest is well understood and can probably be controlled for.
If mdtest hard and mdtest easy as we’ve discussed can be gamed, then any other benchmark can as well.  But if someone “games” to do well with mdtest hard and mdtest easy, then I do think that sets bounds for user expectations of metadata performance.

2. We have had a lot of discussion of moving to object storage because we don’t have a choice. The vendors are addressing the needs of their 95% customer. I don’t think IOR is a fair test of this. It ends up being a test of the mapping from the data structures to objects. For example, using something like PLFS will be a HUGE advantage for “fixing” the IO to be more object oriented no matter the IO API/middleware limitations. In essence, you could cheat trivially.
As I discussed above, PLFS would fail horribly.

3. By doing mdtest and ior separately, we are decoupling the two. Striping issues that hit the metadata server are part of the file creation AND IO performance issues. Do we want to combine these in a more direct test somehow?
Maybe we should do IOR hard, IOR easy, mdtest hard, mdtest easy independently as four measurements.  Then run all four concurrently as a fifth measurement?

4. How much of what we are testing is intended to be the hardware vs. the storage software layer (e.g., Lustre) vs. the middleware (MPI-IO + PLFS vs. ADIOS + BP) vs. IO API (HDF/NetCDF vs. ADIOS vs. POSIX vs. MPI-IO)? Testing at each of these levels makes a lot of sense and have different values to different audiences. I’d argue that all that matters is the top level test since we are trying to support applications. If they do N-1 files, unless the system ALWAYS uses PLFS, it should suffer the stack performance characteristics. Doing something “simple” at a lower layer does not represent what end users care about—what IO performance can I expect? I think IOR can do a lot of it, but it isn’t a complete solution.
I agree that all that matters is the top-level test since we are trying to support apps.  I’d phrase it as ‘we are trying to help apps predict their performance’ as this was a comment made at the BoF by someone whose name I sadly do not know.

5. How do we deal with burst buffers in their various incarnations? Do we make rules about relative sizes of BB and main memory to decide if other storage systems have to be considered? Is there a different metric such as accessible from some external location that determines what we want to benchmark? Is that fair since many systems are being bought with a BB to hide that latency in the general case believing that there are sufficient IOPS and back end bandwidth to drain without slowing applications.
I like 5 minutes of sustained IO.  If the BB is large enough to get super high bandwidth during 5 minutes, then I’m willing to believe its a good storage system.  Sure, someone might build a BB sized for 5 minutes and then not even bother to have a second tier just because they want to win IO 500.  And they’ve done dumb stuff like that for Top 500 too.  We can’t build a perfect benchmark.

But . . . I’m willing to spend some time trying to build a good one.  Although, to be honest, I remain of the opinion that IOR and mdtest are already a good one.  I’ve heard good arguments against them but nothing sufficient to persuade me that we can do any better.  Maybe I just haven’t understood the arguments well enough.

There are tons more things to consider.
Agreed.  So many that attempting to consider them all dooms us to inaction.



From: IO-500 <io-500-bounces@vi4io.org> on behalf of Julian Kunkel <juliankunkel@googlemail.com>
Date: Tuesday, November 22, 2016 at 12:49 AM
To: John Bent <John.Bent@seagategov.com>
Cc: "io-500@vi4io.org" <io-500@vi4io.org>
Subject: [EXTERNAL] Re: [IO-500] Benchmark abstraction

Dear All,
I'm not *against* using IOR but at this stage, I  rather favour a clear separation between
What and why certain metrics are useful to be measured and in a second step 
How they are measured.

This also serves as validation that we do the right thing. I found this always useful when defining a test, and a benchmark is just a performance test for me. The intended purpose helps not only in communication but also prevents unintentional optimizations of systems.

Again I agree that IOR could be the vehicle but I would hope the community firstly agrees on the metrics before there might be detailed discussions about the tool.


Am 21.11.2016 10:37 nachm. schrieb "John Bent" <John.Bent@seagategov.com>:
Thanks Sarp!  Some comments in-line. 
On Nov 21, 2016, at 2:22 PM, Oral, H. Sarp <oralhs@ornl.gov> wrote:

Well, I agree with John that trying to define a new and all around benchmark is highly difficult. We tried that (and looked a few other benchmarks at the time) and failed. No need to repeat the same mistakes, I think.

And I also agree that the benchmarks need to be simple and easy to run and representative of realistic scenarios.

Rather than limiting to two IOR instances, we can perhaps increase them slightly to cover more I/O workloads with IOR, if needed.

By the way, we already have an IOR version that we integrated with ADIOS. We can share it with the community. And IOR already supports HDF5, and MPIIO. Between POSIX and these mid level libs, I think IOR covers a majority of the use cases. The trick is coming up with good, canned command line option sets for IOR covering various I/O workloads.

There is really nothing else on measuring the mdtest today as far as I know.

In terms of mdtest, when I speak of it, I have to admit that I’m speaking of a theoretical future mdtest which does not yet exist.  IOR is beautifully engineered with a fantastic plug-in feature as you mention.  The mdtest I’m envisioning is taking mdtest.c from it’s current github and moving it into the IOR github and rewriting it to replace the POSIX calls into calls to this IOR plug-in interface.  The plug-in interface is already almost a superset of what mdtest needs.  I think only ‘stat’ needs to be added.  That way, when people add new plug-ins to IOR, they will simultaneously add them to mdtest.  Also, for our benchmark, they’d simply pull and ‘make’ from a single repository.
Any volunteers?  :)
Here’s what I believe to be the most recently maintained repositories:
I have to admit that I have not yet looked at md-real-io to do a comparison.  (sorry Julian, it is on my TODO list…)

So, we are on board.

Fantastic.  Our survey is now 1% complete.  :)



Sarp Oral, PhD

National Center for Computational Sciences
Oak Ridge National Laboratory

On 11/20/16, 12:33 PM, "IO-500 on behalf of Julian Kunkel" <io-500-bounces@vi4io.orgon behalf of juliankunkel@googlemail.com> wrote:

   Dear John,
   I would definitely not go with mdtest. That one can be well optimized by read ahead / sync. Also it is POSIX only.
   Note that for overcoming the caching problem, I wrote the md-real-io benchmark that shares many things with mdtest.
   I would wait for the community feedback and not ignore that concepts such as ADIOS may not necessarily fit as IOR back ends and rather to with abstract definitions first.


   Am 20.11.2016 6:12 nachm. schrieb "John Bent" <John.Bent@seagategov.com>:

   To attempt defining the perfect IO benchmark is Quixotic.  Those who dislike IO500 will always dislike IO500 regardless of what the specific benchmark is.  Those who like the idea will accept an imperfect benchmark.

   Therefore, I suggest we move forward with the straw person proposal: IOR hard, IOR easy, mdtest hard, mdtest easy.  

   * Average IOR hard and IOR easy.  Average mdtest hard and mdtest easy.  
   * Their product determines the winner.  
   * Don’t report the product since it’s a meaningless unit; report the averages.
   * e.g. The winner of IO500 is TaihuLight with a score of 250 GB/s and 300K IOPs.

   Unless a proposal is strictly much better than IOR hard, IOR easy, mdtest hard, mdtest easy, I don’t think we should consider it.  The beauty of IOR hard, IOR easy, mdtest hard, mdtest easy is that they are well-understood, well-accepted benchmarks, that
    are trivial to download and compile, and whose results are immediately understandable.  Every RFP in the world uses them.  The one problem is they need a pithier name than “IOR hard, IOR easy, mdtest hard, mdtest easy”...

   My suggestion is to poll the top 100 of the 
   top500.org <http://top500.org> and ask them this:

   "If we were to do an IO500, and our benchmark was IOR hard, IOR easy, mdtest hard, mdtest easy, would you participate?  If not, would you participate with a different benchmark?”

   If the bulk of the answers are “yes,” then we just figure out how to organize and administer this thing.
   If the bulk of the answers are “no,” then we give up and do something else. 
   If the bulk of the answers are “no, yes,” then we need to find a new benchmark.



   On Nov 20, 2016, at 7:11 AM, Julian Kunkel <juliankunkel@googlemail.com> wrote:

   Dear all,
   based on our discussion during the BoF at SC, we could focus on the
   access pattern(s) of interest first. Later we can define which
   benchmarks (such as IOR) could implement these patterns (e.g., how to
   call existing benchmarks).

   This strategy gives other I/O paradigms the option to create a
   benchmark with that pattern that fits their I/O paradigm/architecture.

   Here is a draft of one that is probably not too difficult to discuss:
   Goal: IOmax: Sustained performance for well-formed I/O

   The benchmark shall determine the best sustained I/O behavior without
   in-memory caching and I/O variability. A set of real applications that
   are highly optimized should be able to show the described access

   Use case: A large data structure is distributed across N
   threads/processes; a time series of this data structured shall be
   stored/retrieved efficiently. (This could be a checkpoint.)

   Processing steps:
   S0) Each thread allocates and initializes a large consecutive memory
   region of size S with a random (but well defined) pattern
   S1) Repeat T times: Each process persists/reads its data to/from the
   storage. Ech iteration  is protected with a global barrier and the
   runtime is measured
   S2) Compute the throughput (as IOmax) by dividing the total accessed
   data volume (N*S) by the maximum observed runtime for any single
   iteration in step S1

   R1) The data of each thread and timestep must be stored individually
   and cannot be overwritten during a benchmark run
   R2) It must be ensured that the time includes all processes needed to
   persist all data in volatile memory (for writes) and that prior
   startup of reads no data is cached in any volatile memory
   R3) A valid result must verify that read returns the expected (random) data
   R4) N, T and S can be set arbitrarily. T must be >= 3. The benchmark
   shall be repeated several times

   Reported metrics:
   * IOmax
   * Working set size W: N*T*S

   IO-500 mailing list

   confidential and proprietary material for the sole use of
   the intended recipient. Any review or distribution by others
   is strictly prohibited. If you are not the intended recipient
   please contact the sender and delete all copies.


confidential and proprietary material for the sole use of
the intended recipient. Any review or distribution by others
is strictly prohibited. If you are not the intended recipient
please contact the sender and delete all copies.

confidential and proprietary material for the sole use of
the intended recipient. Any review or distribution by others
is strictly prohibited. If you are not the intended recipient
please contact the sender and delete all copies.
IO-500 mailing list

confidential and proprietary material for the sole use of
the intended recipient. Any review or distribution by others
is strictly prohibited. If you are not the intended recipient
please contact the sender and delete all copies.