[
https://issues.apache.org/jira/browse/MESOS-7385?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jay Guo updated MESOS-7385:
---------------------------
Description:
Mesos currently implements naive H-DRF algorithm, as described in [h-drf
paper|https://people.eecs.berkeley.edu/~alig/papers/h-drf.pdf], which may incur
starvation due to `dovetailing`. Essentially, following test should pass:
{code}
TEST_F(HierarchicalAllocatorTest, Starvation)
{
Clock::pause();
initialize();
const string ROLE1 = "a";
const string ROLE2 = "b/c";
const string ROLE3 = "b/d";
FrameworkInfo framework1 = createFrameworkInfo({ROLE1});
allocator->addFramework(framework1.id(), framework1, {}, true);
SlaveInfo agent1 = createSlaveInfo("cpus:1");
allocator->addSlave(
agent1.id(),
agent1,
AGENT_CAPABILITIES(),
None(),
agent1.resources(),
{});
// `framework1` will be offered all of the resources on `agent1`.
{
Allocation expected = Allocation(
framework1.id(),
{{ROLE1, {{agent1.id(), agent1.resources()}}}});
AWAIT_EXPECT_EQ(expected, allocations.get());
}
// Create `framework2` in the child role.
FrameworkInfo framework2 = createFrameworkInfo({ROLE2});
allocator->addFramework(framework2.id(), framework2, {}, true);
SlaveInfo agent2 = createSlaveInfo("mem:32");
allocator->addSlave(
agent2.id(),
agent2,
AGENT_CAPABILITIES(),
None(),
agent2.resources(),
{});
{
Allocation expected = Allocation(
framework2.id(),
{{ROLE2, {{agent2.id(), agent2.resources()}}}});
AWAIT_EXPECT_EQ(expected, allocations.get());
}
// Create `framework3` in the child role.
FrameworkInfo framework3 = createFrameworkInfo({ROLE3});
allocator->addFramework(framework3.id(), framework3, {}, true);
SlaveInfo agent3 = createSlaveInfo("cpus:1");
allocator->addSlave(
agent3.id(),
agent3,
AGENT_CAPABILITIES(),
None(),
agent3.resources(),
{});
// Current fair share is:
// - `framework1`: 50% (1/2 cpus)
// - `framework2`: 100% (32/32 mem)
// - `framework3`: 0% (0/2 cpus)
// So `framework3` should be offered all of the resources on `agent3`.
// However, `framework3` is punished due to naive h-drf implementation,
// where fair share of parent role `b` has fair share of 100%, which
// leads to starvation.
{
Allocation expected = Allocation(
framework3.id(),
{{ROLE3, {{agent3.id(), agent3.resources()}}}});
AWAIT_EXPECT_EQ(expected, allocations.get()); // It fails!
}
}
{code}
This JIRA is created to make sure this behavior is captured and will be
addressed in the future. Note that it affects current implementation without
hierarchical role as well.
was:
Mesos currently implements naive H-DRF algorithm, as described in [h-drf
paper|https://people.eecs.berkeley.edu/~alig/papers/h-drf.pdf], which may incur
starvation due to `dovetailing`. Essentially, following test should pass:
{{code}}
TEST_F(HierarchicalAllocatorTest, Starvation)
{
Clock::pause();
initialize();
const string ROLE1 = "a";
const string ROLE2 = "b/c";
const string ROLE3 = "b/d";
FrameworkInfo framework1 = createFrameworkInfo({ROLE1});
allocator->addFramework(framework1.id(), framework1, {}, true);
SlaveInfo agent1 = createSlaveInfo("cpus:1");
allocator->addSlave(
agent1.id(),
agent1,
AGENT_CAPABILITIES(),
None(),
agent1.resources(),
{});
// `framework1` will be offered all of the resources on `agent1`.
{
Allocation expected = Allocation(
framework1.id(),
{{ROLE1, {{agent1.id(), agent1.resources()}}}});
AWAIT_EXPECT_EQ(expected, allocations.get());
}
// Create `framework2` in the child role.
FrameworkInfo framework2 = createFrameworkInfo({ROLE2});
allocator->addFramework(framework2.id(), framework2, {}, true);
SlaveInfo agent2 = createSlaveInfo("mem:32");
allocator->addSlave(
agent2.id(),
agent2,
AGENT_CAPABILITIES(),
None(),
agent2.resources(),
{});
{
Allocation expected = Allocation(
framework2.id(),
{{ROLE2, {{agent2.id(), agent2.resources()}}}});
AWAIT_EXPECT_EQ(expected, allocations.get());
}
// Create `framework3` in the child role.
FrameworkInfo framework3 = createFrameworkInfo({ROLE3});
allocator->addFramework(framework3.id(), framework3, {}, true);
SlaveInfo agent3 = createSlaveInfo("cpus:1");
allocator->addSlave(
agent3.id(),
agent3,
AGENT_CAPABILITIES(),
None(),
agent3.resources(),
{});
// Current fair share is:
// - `framework1`: 50% (1/2 cpus)
// - `framework2`: 100% (32/32 mem)
// - `framework3`: 0% (0/2 cpus)
// So `framework3` should be offered all of the resources on `agent3`.
// However, `framework3` is punished due to naive h-drf implementation,
// where fair share of parent role `b` has fair share of 100%, which
// leads to starvation.
{
Allocation expected = Allocation(
framework3.id(),
{{ROLE3, {{agent3.id(), agent3.resources()}}}});
AWAIT_EXPECT_EQ(expected, allocations.get()); // It fails!
}
}
{{code}}
This JIRA is created to make sure this behavior is captured and will be
addressed in the future. Note that it affects current implementation without
hierarchical role as well.
> Framework should not starve due to `dovetailing` in naive H-DRF
> implementation.
> -------------------------------------------------------------------------------
>
> Key: MESOS-7385
> URL: https://issues.apache.org/jira/browse/MESOS-7385
> Project: Mesos
> Issue Type: Bug
> Components: master
> Reporter: Jay Guo
>
> Mesos currently implements naive H-DRF algorithm, as described in [h-drf
> paper|https://people.eecs.berkeley.edu/~alig/papers/h-drf.pdf], which may
> incur starvation due to `dovetailing`. Essentially, following test should
> pass:
> {code}
> TEST_F(HierarchicalAllocatorTest, Starvation)
> {
> Clock::pause();
> initialize();
> const string ROLE1 = "a";
> const string ROLE2 = "b/c";
> const string ROLE3 = "b/d";
> FrameworkInfo framework1 = createFrameworkInfo({ROLE1});
> allocator->addFramework(framework1.id(), framework1, {}, true);
> SlaveInfo agent1 = createSlaveInfo("cpus:1");
> allocator->addSlave(
> agent1.id(),
> agent1,
> AGENT_CAPABILITIES(),
> None(),
> agent1.resources(),
> {});
> // `framework1` will be offered all of the resources on `agent1`.
> {
> Allocation expected = Allocation(
> framework1.id(),
> {{ROLE1, {{agent1.id(), agent1.resources()}}}});
> AWAIT_EXPECT_EQ(expected, allocations.get());
> }
> // Create `framework2` in the child role.
> FrameworkInfo framework2 = createFrameworkInfo({ROLE2});
> allocator->addFramework(framework2.id(), framework2, {}, true);
> SlaveInfo agent2 = createSlaveInfo("mem:32");
> allocator->addSlave(
> agent2.id(),
> agent2,
> AGENT_CAPABILITIES(),
> None(),
> agent2.resources(),
> {});
> {
> Allocation expected = Allocation(
> framework2.id(),
> {{ROLE2, {{agent2.id(), agent2.resources()}}}});
> AWAIT_EXPECT_EQ(expected, allocations.get());
> }
> // Create `framework3` in the child role.
> FrameworkInfo framework3 = createFrameworkInfo({ROLE3});
> allocator->addFramework(framework3.id(), framework3, {}, true);
> SlaveInfo agent3 = createSlaveInfo("cpus:1");
> allocator->addSlave(
> agent3.id(),
> agent3,
> AGENT_CAPABILITIES(),
> None(),
> agent3.resources(),
> {});
> // Current fair share is:
> // - `framework1`: 50% (1/2 cpus)
> // - `framework2`: 100% (32/32 mem)
> // - `framework3`: 0% (0/2 cpus)
> // So `framework3` should be offered all of the resources on `agent3`.
> // However, `framework3` is punished due to naive h-drf implementation,
> // where fair share of parent role `b` has fair share of 100%, which
> // leads to starvation.
> {
> Allocation expected = Allocation(
> framework3.id(),
> {{ROLE3, {{agent3.id(), agent3.resources()}}}});
> AWAIT_EXPECT_EQ(expected, allocations.get()); // It fails!
> }
> }
> {code}
> This JIRA is created to make sure this behavior is captured and will be
> addressed in the future. Note that it affects current implementation without
> hierarchical role as well.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
