A benchmark of Fedora-4 perfomance

[jcoyne]

time = Benchmark.measure do
  (1..1000).each { GenericFile.create! { |f| f.apply_depositor_metadata('jcoyne') } }
end
=> #<Benchmark::Tms:0x007fb240810548 @label="", @real=631.307004, @cstime=0.0, @cutime=0.0, @stime=32.4, @utime=261.5, @total=293.9>
puts time
261.500000  32.400000 293.900000 (631.307004)

That's an ingest rate of 95 records per minute. Interestingly the CPU was not maxed out, most of the time was spent waiting on IO (fedora/solr)

[jcoyne]

Here's what happens in each create:

• Post object to Fedora
• Check for (HEAD) child "content" (3x)
• Check for (HEAD) child "thumbnail" (3x)
• Check for (HEAD) child "full_text" (3x)
• Check for (HEAD) the object
• Retrieve (GET) the object
• Check for (HEAD) child "characterization"
• Check for (HEAD) child "full_text"
• Query Solr for collection members (4x)
• Write the object to solr
• POST ACL to Fedora
• Check for (HEAD) the ACL
• Retrieve (GET) the ACL
• Write the object to solr (with ACL data)

See a full log here: https://gist.github.com/jcoyne/b46309ce0a81fa6e2782

There are some clear opportunities for reducing the duplicate (3x and 4x) requests. Since the object has assertions about it's children. We may be able to remove the HEAD requests when children aren't present too.

[carolyncole]

@jcoyne Does look like we are hitting fedora too much and could use some caching. Are we to the point where we want to tackle this, or is it something for the future?

[jcoyne]

@Cam156 I think now is a good time to look at improving performance if we don't have anything critical to work on. We are hitting fedora just a bit too much and we could perhaps shave 100ms off each create. One problem is that we have to Push to Solr twice because of the logic that validates an object has a depositor and that the depositor should have edit access. I'm wondering if there's any way around that.

[hectorcorrea]

@jcoyne what do we get from the HEAD calls? Do we really need them? (They are in LDP, right?)

[jcoyne]

@hectorcorrea you get an acknowledgement that the resource exists without having to load the resource itself. They are faster particularly for large resources.

[jcoyne]

@hectorcorrea additionally you get all the headers like Content-Disposition, Content-Type, Content-Length, etc.

[jcoyne]

samvera/ldp#34 removes the duplicate 404s.

[jcoyne]

samvera-deprecated/hydra-collections#50 halves the number of Solr calls required to index the collections.

[jcoyne]

samvera/active_fedora#645 further reduces calls to Solr when objects are not members of any collections.

[jcoyne]

irb(main):006:0> time = Benchmark.measure do
irb(main):007:1*   (1..1000).each { GenericFile.create! { |f| f.apply_depositor_metadata('jcoyne') } }
irb(main):008:1> end
=> #<Benchmark::Tms:0x007fcf76fd7440 @label="", @real=410.517791, @cstime=0.0, @cutime=0.0, @stime=26.590000000000003, @utime=235.13, @total=261.72>
irb(main):009:0> puts time
235.130000  26.590000 261.720000 (410.517791)
=> nil

Yay, now we're up to 146 per minute..

[jcoyne]

Compare to Fedora 3:

> puts time
490.660000  16.990000 507.650000 (737.432756)

[hectorcorrea]

👍

[mjgiarlo]

How many per minute is F3, @jcoyne?

[rotated8]

Roughly 81 per minute, I believe.