Qian Zhang created MESOS-3244:
---------------------------------
Summary: Enhancement for the allocator() method in the default DRF
allocator
Key: MESOS-3244
URL: https://issues.apache.org/jira/browse/MESOS-3244
Project: Mesos
Issue Type: Improvement
Components: allocation
Reporter: Qian Zhang
Assignee: Qian Zhang
Priority: Minor
In the method *HierarchicalAllocatorProcess::allocate(const hashset<SlaveID>&
slaveIds)*, I see the following code:
{code}
template <class RoleSorter, class FrameworkSorter>
void
HierarchicalAllocatorProcess<RoleSorter, FrameworkSorter>::allocate(
const hashset<SlaveID>& slaveIds_)
{
...
foreach (const SlaveID& slaveId, slaveIds) {
...
foreach (const std::string& role, roleSorter->sort()) {
foreach (const std::string& frameworkId_, frameworkSorters[role]->sort())
{
FrameworkID frameworkId;
frameworkId.set_value(frameworkId_);
// Calculate the currently available resources on the slave.
Resources available = slaves[slaveId].total - slaves[slaveId].allocated;
// NOTE: Currently, frameworks are allowed to have '*' role.
// Calling reserved('*') returns an empty Resources object.
Resources resources = available.unreserved() + available.reserved(role);
...
}
}
}
}
{code}
My question is why the following two lines are put in the innermost loop (the
3rd level foreach for framework).
{code}
Resources available = slaves[slaveId].total - slaves[slaveId].allocated;
Resources resources = available.unreserved() + available.reserved(role);
{code}
I think these two line has nothing to do with framework (they calculate
available resources ONLY based on slave and role), but since they are in the
innermost loop, they will be executed for EVERY framework of a role (imagine we
have thousand frameworks for a role), but I think they should be only executed
once for a role.
So I would suggest to move these two lines up, they should be in the 2nd level
loop (the foreach loop for role), and in the 3rd level loop, for each framework
of the role, we should directly use the result of these two lines rather than
do the calculation for each framework.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
