> On March 17, 2017, 1:33 a.m., Benjamin Mahler wrote:
> > Hm.. this seems to introduce a performance regression in the case where there are
many inactive roles in the system? I assume it's difficult to avoid in the hierarchical case?
Otherwise, this looks good.
>
> Neil Conway wrote:
> True -- if there are many inactive clients, we'll now calculate their share and include
them in the `std::set`, whereas we wouldn't do so before (conversely, activating and deactivating
clients is a lot faster than it used to be). My guess is that the bottlenecks in sorter performance
likely lie elsewhere (e.g., updating resources, generating the entire `vector` in `sort()`
when the allocator might only need the first few entries).
>
> We could avoid this overhead by skipping `calculateShare` for inactive clients, and
then dirtying the whole sorter when a client is activated. My guess is that this isn't a net
win (if you have a lot of inactive clients, it seems unfortunate to dirty the entire sorter
whenever a client becomes active), but I can do some benchmarks if you think this case is
important.
Another alternative to avoid dirtying the whole sorter when a client is activated is to a
per-client dirty bit, or `Option<double> share` where if set, it is accurate (and we
can unset to induce the lazy calculation). `dirty = true` would be equivalent to clearing
all of the clients' shares. Thinking about it, the `Option<double>` seems simpler to
understand as well (no way for a stale value to be there).
- Benjamin
-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/57564/#review169245
-----------------------------------------------------------
On March 13, 2017, 6:04 p.m., Neil Conway wrote:
>
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/57564/
> -----------------------------------------------------------
>
> (Updated March 13, 2017, 6:04 p.m.)
>
>
> Review request for mesos, Benjamin Bannier, Benjamin Mahler, and Michael Park.
>
>
> Repository: mesos
>
>
> Description
> -------
>
> DRFSorter previously removed inactive clients from the `clients`
> collection, and then re-added clients when they were reactivated. This
> resulted in resetting the allocation count for the client, which is
> unfortunate. This scheme would also be more difficult to adapt to
> hierarchical sorting.
>
> This commit changes DRFSorter to continue to store inactive clients in
> the `clients`; inactive clients are indicated by a new field in the
> `Client` struct, and are omitted from the return value of
> `DRFSorter::sort`.
>
>
> Diffs
> -----
>
> src/master/allocator/sorter/drf/sorter.hpp 76329220e1115c1de7810fb69b943c78c078be59
> src/master/allocator/sorter/drf/sorter.cpp ed54680cecb637931fc344fbcf8fd3b14cc24295
> src/tests/sorter_tests.cpp ec0636beb936d46a253d19322f2157abe95156b6
>
>
> Diff: https://reviews.apache.org/r/57564/diff/1/
>
>
> Testing
> -------
>
> `make check`
>
>
> Thanks,
>
> Neil Conway
>
>
|