mesos-reviews mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Till Toenshoff <toensh...@me.com>
Subject Review Request 67228: Added allocator benchmark test for various simulations.
Date Sat, 19 May 2018 00:05:59 GMT

-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/67228/
-----------------------------------------------------------

Review request for mesos, Benjamin Mahler, Greg Mann, Kapil Arya, and Vinod Kone.


Repository: mesos


Description
-------

This benchmark measures the performannce of the DRF allocator.

* Accept rate (`min_accept_rate`, `max_accept_rate`, and
  `accept_rate_divisor`): Each framework is initialized with a given
  offer accept-rate. This rate determines the number of offers
  accepted out of the total received.
  The first framework is launched with `max_accept_rate`. Each
  subsequent framework is launched with an accept rate of
  (previous framework's accept-rate / `accept_rate_divisor`). This
  rate will never be below `min_accept_rate`.
  For example, with a `max_accept_rate` of 1 and `accept_rate_divisor`
  of 2, the framework accept rates will be 1, 0.5, 0.25, 0.125, and
  so on.


* Terminating condition(s): the benchmark stops if any of the
  following condition is met:
    * Number of offers generated reaches `max_offers`.
    * Number of total tasks launched reaches `max_tasks`.
    * Any single framework launches `max_framework_tasks` tasks.


* Offer decline timeout (`decline_timeout` and
  `partial_decline_timeout`): When declining an entire offer, the
  offer-filter timeout is set to `decline_timeout` for a given slave.
  If an offer is partially accepted, the offer-filter timeout is set
  to `partial_decline_timeout`. If the two timeouts are the same,
  framework starvation is observed for framework with the highest
  accept rate. To avoid such starvation, one can try setting
  `decline_timeout` to be significantly higher (e.g., 10x) than
  `partial_decline_timeout`.


Diffs
-----

  src/Makefile.am c08ac6e2f5deec4d05f59f71ff6c51382f216708 
  src/tests/master_drf_benchmarks.cpp PRE-CREATION 


Diff: https://reviews.apache.org/r/67228/diff/1/


Testing
-------

WIP - still validating results - WIP


Thanks,

Till Toenshoff


Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message