Performance Regressions

[arnavs]

The odu lecture, for ex., looks like it takes 3X as long to run now than it did with Julia 0.6. What we need to do is:

• [x] Cache/store the build information/runtimes for Julia 0.6.
• [ ] Identify affected lectures.
• [ ] Fix them.
• [ ] (long-term) Add some kind of benchmarking to the build process so we can find these automatically from Travis.

[arnavs]

v0.6 data can be found as a JSON here

[arnavs]

@XiaojunGuan This might be a good one to chip away at.

As a first step, could you look at the v0.6 data in the pastebin, and the v1.0 data at https://lectures.quantecon.org/status.html, and identify which lectures have slowed down?

Edit: Just to confirm, we only care about the Julia times, and things on the order of a few seconds here or there aren't too worrisome.

[XiaojunGuan]

Here is a list of lectures that are being affected:

• [ ] dyn_stack
• [ ] egm_policy_iter
• [ ] mccall_model
• [ ] need_for_speed
• [ ] smoothing

[jlperla]

We need to be careful about precompiling time. The test should only compare running the notebooks the second time (in both circumstances)

[arnavs]

Thanks @XiaojunGuan.

[arnavs]

And that's a good point @jlperla --- I'll confirm with Matt.

[jlperla]

No, I mean when manually running it... It is already too slow to run twice until we can parallelize the build