Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[WIP] Add co-scheduling functionality for GPUs using MPS #1317

Draft
wants to merge 4 commits into
base: master
Choose a base branch
from
Draft

[WIP] Add co-scheduling functionality for GPUs using MPS #1317

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
587f2f8
allow non-exclusive slots
kulnaman Dec 5, 2024
7963032
enable MPS. This includes:
kulnaman Dec 5, 2024
5672c1c
MPS Add distributive GPU allocation
kulnaman Dec 5, 2024
f90cf68
Merge branch 'flux-framework:master' into mps_cosched
kulnaman Dec 5, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 31 additions & 0 deletions mps_data/job_1.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,31 @@
version: 9999
resources:
- type: node
count: 1
with:
- type: socket
count: 1
with:
- type: slot
label: default
exclusive: false
count: 1
with:
- type: core
count: 1
exclusive: true
- type: gpu
count: 1
exclusive: false

# a comment
attributes:
system:
duration: 120
cosched: true
c_r: true
tasks:
- command: [ "app" ]
slot: default
count:
per_slot: 1
31 changes: 31 additions & 0 deletions mps_data/job_1N.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,31 @@
version: 9999
resources:
- type: node
count: 1
with:
- type: socket
count: 1
with:
- type: core
count: 2
- type: gpu
count: 1
with:
- type: slot
label: default
count: 1
with:
- type: gpu_mps
count: 1

# a comment
attributes:
system:
duration: 120
cosched: true
c_r: true
tasks:
- command: [ "app" ]
slot: default
count:
per_slot: 1
35 changes: 35 additions & 0 deletions mps_data/job_2.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,35 @@
version: 9999
resources:
- type: node
count: 1
with:
- type: socket
count: 1
with:
- type: slot
label: default
exclusive: false
count: 1
with:
- type: core
count: 2
exclusive: true
- type: gpu
count: 2
exclusive: false
with:
- type: gpu_mps
count: 1
exclusive: true

# a comment
attributes:
system:
duration: 120
cosched: true
c_r: true
tasks:
- command: [ "app" ]
slot: default
count:
per_slot: 1
27 changes: 27 additions & 0 deletions mps_data/job_2N.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,27 @@
version: 9999
resources:
- type: node
count: 4
with:
- type: socket
count: 1
with:
- type: slot
label: default
count: 1
with:
- type: core
count: 4
- type: gpu
count: 4
# a comment
attributes:
system:
duration: 720
cosched: true
c_r: true
tasks:
- command: [ "app" ]
slot: default
count:
per_slot: 1
27 changes: 27 additions & 0 deletions mps_data/job_4N.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,27 @@
version: 9999
resources:
- type: node
count: 4
with:
- type: socket
count: 1
with:
- type: slot
label: default
count: 1
with:
- type: core
count: 2
- type: gpu
count: 2
# a comment
attributes:
system:
duration: 240
cosched: true
c_r: true
tasks:
- command: [ "app" ]
slot: default
count:
per_slot: 1
138 changes: 138 additions & 0 deletions mps_data/small.graphml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,138 @@
<?xml version="1.0" encoding="UTF-8"?>

<!-- 1 subsystem, cluster[1]->rack[1]->node[8]->socket[2]->core[18] -->

<!-- Memory pool is modeled as 4 x 32GB -->


<graphml xmlns="http://graphml.graphdrawing.org/xmlns">
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://graphml.graphdrawing.org/xmlns
http://graphml.graphdrawing.org/xmlns/1.1/graphml.xsd">

<!-- resource pool vertex generation spec attributes -->
<key id="root" for="node" attr.name="root" attr.type="int">
<default>0</default>
</key>
<key id="type" for="node" attr.name="type" attr.type="string"/>
<key id="basename" for="node" attr.name="basename" attr.type="string"/>
<key id="unit" for="node" attr.name="unit" attr.type="string"/>
<key id="size" for="node" attr.name="size" attr.type="long">
<default>1</default>
</key>
<key id="subsystem" for="node" attr.name="subsystem" attr.type="string">
<default>containment</default>
</key>

<!-- resource relationship generation attributes -->
<key id="e_subsystem" for="edge" attr.name="e_subsystem" attr.type="string">
<default>containment</default>
</key>
<key id="relation" for="edge" attr.name="relation" attr.type="string">
<default>contains</default>
</key>
<key id="rrelation" for="edge" attr.name="rrelation" attr.type="string">
<default>in</default>
</key>

<!-- id generation method -->
<key id="id_scope" for="edge" attr.name="id_scope" attr.type="int">
<default>0</default>
</key>
<key id="id_start" for="edge" attr.name="id_start" attr.type="int">
<default>0</default>
</key>
<key id="id_stride" for="edge" attr.name="id_stride" attr.type="int">
<default>1</default>
</key>

<!-- resource gen method: multiply or associate-in -->
<key id="gen_method" for="edge" attr.name="gen_method" attr.type="string">
<default>MULTIPLY</default>
</key>
<!-- argument (scaling factor) for multiply method -->
<key id="multi_scale" for="edge" attr.name="multi_scale" attr.type="int">
<default>1</default>
</key>
<!-- 3 arguments for associate-in method -->
<key id="as_tgt_subsystem" for="edge" attr.name="as_tgt_subsystem"
attr.type="string">
<default>containment</default>
</key>
<key id="as_tgt_uplvl" for="edge" attr.name="as_tgt_uplvl" attr.type="int">
<default>1</default>
</key>
<key id="as_src_uplvl" for="edge" attr.name="as_src_uplvl" attr.type="int">
<default>1</default>
</key>


<!-- generation recipe for the small cluster -->
<graph id="small" edgedefault="directed">

<!-- containment subsystem generation recipe -->
<node id="cluster">
<data key="root">1</data>
<data key="type">cluster</data>
<data key="basename">small</data>
</node>
<node id="rack">
<data key="type">rack</data>
<data key="basename">rack</data>
</node>
<node id="node">
<data key="type">node</data>
<data key="basename">node</data>
</node>
<node id="socket">
<data key="type">socket</data>
<data key="basename">socket</data>
</node>
<node id="core">
<data key="type">core</data>
<data key="basename">core</data>
</node>
<node id="gpu">
<data key="type">gpu</data>
<data key="basename">gpu</data>
</node>
<node id="gpu_mps">
<data key="type">gpu_mps</data>
<data key="basename">gpu_mps</data>
</node>
<node id="memory">
<data key="type">memory</data>
<data key="basename">memory</data>
<data key="size">32</data>
<data key="unit">GB</data>
</node>

<edge id="cluster2rack" source="cluster" target="rack">
<data key="multi_scale">1</data>
</edge>
<edge id="rack2node" source="rack" target="node">
<data key="id_scope">1</data>
<data key="multi_scale">2</data>
</edge>
<edge id="node2socket" source="node" target="socket">
<data key="multi_scale">1</data>
</edge>
<edge id="socket2core" source="socket" target="core">
<data key="id_scope">1</data>
<data key="multi_scale">8</data>
</edge>
<edge id="socket2gpu" source="socket" target="gpu">
<data key="id_scope">1</data>
<data key="multi_scale">2</data>
</edge>
<edge id="gpu2gpu_mps" source="gpu" target="gpu_mps">
<data key="id_scope">1</data>
<data key="multi_scale">2</data>
</edge>
<edge id="socket2memory" source="socket" target="memory">
<data key="id_scope">1</data>
<data key="multi_scale">4</data>
</edge>
</graph>
</graphml>

2 changes: 2 additions & 0 deletions resource/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@ set(RESOURCE_HEADERS
policies/dfu_match_multilevel_id_impl.hpp
policies/dfu_match_locality.hpp
policies/dfu_match_var_aware.hpp
policies/dfu_match_cosched_aware.hpp
policies/dfu_match_policy_factory.hpp
jobinfo/jobinfo.hpp
schema/resource_graph.hpp
Expand Down Expand Up @@ -52,6 +53,7 @@ set(RESOURCE_HEADERS
add_library(resource STATIC
policies/dfu_match_locality.cpp
policies/dfu_match_var_aware.cpp
policies/dfu_match_cosched_aware.cpp
policies/dfu_match_policy_factory.cpp
jobinfo/jobinfo.cpp
schema/resource_data.cpp
Expand Down
21 changes: 20 additions & 1 deletion resource/evaluators/scoring_api.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -112,6 +112,10 @@ void scoring_api_t::merge (const scoring_api_t &o)
}
}

void scoring_api_t::add_element_to_child_score (int score)
{
children_score_list.push_back (score);
}
void scoring_api_t::resrc_types (subsystem_t s, std::vector<resource_type_t> &v)
{
for (auto &kv : m_ssys_map[s])
Expand Down Expand Up @@ -139,7 +143,22 @@ void scoring_api_t::set_avail (unsigned int avail)
{
m_avail = avail;
}

void scoring_api_t::set_children_score_vector (const std::vector<int64_t> source)
{
children_score_list = source;
}
std::vector<int64_t> scoring_api_t::get_children_score_vector ()
{
return children_score_list;
}
int64_t scoring_api_t::get_children_avearge_score ()
{
if (children_score_list.size () == 0)
return 0;
return static_cast<int64_t> (
std::accumulate (children_score_list.begin (), children_score_list.end (), 0LL)
/ children_score_list.size ());
}
bool scoring_api_t::is_contained (subsystem_t s, resource_type_t const &r)
{
return m_ssys_map[s].contains (r);
Expand Down
6 changes: 6 additions & 0 deletions resource/evaluators/scoring_api.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,6 +56,11 @@ class scoring_api_t {
void set_overall_score (int64_t overall);
unsigned int avail ();
void set_avail (unsigned int avail);
// Add scores to a vector that keeps the children score. Needed to show preferences above slot also
void add_element_to_child_score(int score);
std::vector<int64_t> get_children_score_vector();
void set_children_score_vector(const std::vector<int64_t> source);
int64_t get_children_avearge_score();
bool is_contained (subsystem_t s, resource_type_t const &r);

template<class compare_op = fold::greater, class binary_op = fold::plus>
Expand Down Expand Up @@ -99,6 +104,7 @@ class scoring_api_t {

private:
intern::interned_key_vec<subsystem_t, std::map<resource_type_t, detail::evals_t>> m_ssys_map;
std::vector<int64_t> children_score_list;
bool m_hier_constrain_now = false;
int64_t m_overall_score = -1;
unsigned int m_avail = 0;
Expand Down
Loading
Loading