Thread Pinning¶
AFF3CT-core
enables to select on which process units (PUs) the threads are
effectively run. This is called thread pinning and it can significantly
benefit to the performance, especially on modern heterogeneous architectures.
To do so, the runtime relies on the
hwloc
library.
Warning
To use thread pinning, hwloc
library has to be installed on the system and
AFF3CT-core
needs to be compiled with the AFF3CT_CORE_HWLOC
preprocessor
definition. It can simply be achieved using the following CMake option:
AFF3CT-core
is not linked with the hwloc
library, then the thread
pinning interface will have no effect and the threads will not be pinned.
Info
Thread pinning relies the OS. The later needs to expose appropriated system calls. While Linux and Windows provide these syscalls, macOS does not... Thus, thread pinning will have no effect on macOS :-(.
Portable Hardware Locality¶
Portable Hardware Locality (hwloc
in short) is a library which provides a
portable abstraction of the hierarchical topology of modern
architectures (see the figure below).
hwloc
gives the ability to pin threads over various level of hierarchy
represented by a tree structure. The deepest/lowest nodes (the leaves) are the
PUs while higher nodes represent sets of PUs that are physically close. For
instance, a PUs set can share the same UMA node (in the case of a NUMA
architecture), the same LLC or the same package.
In the Orange Pi 5 SBC, if we pin a thread on the Package L#0
, it will run
over the following set of PUs: PU L#0
, PU L#1
, PU L#2
and PU L#3
.
Thus, the pinned thread can move in the selected hwloc
node during the
execution and it is up to the OS to schedule the thread on the selected PUs
set.
Warning
The indexes given by hwloc
can be different from those given by the OS:
they are logical indexes that express the real locality. Consequently, in
AFF3CT-core
, it is important to use hwloc
logical indexes. The
hwloc-ls
command gives an overview of the current topology with these
logical indexes.
Sequence & Pipeline¶
In AFF3CT-core
, thread pinning can be set in runtime::Sequence
and
runtime::Pipeline
class constructors. In both cases, there is a dedicated
argument of std::string
type named sequence_pinning_policy
for
runtime::Sequence
or pipeline_pinning_policy
for runtime::Pipeline
.
Info
For NUMA architectures, it is important to specify thread pinning at the
construction of the runtime::Sequence
/runtime::Pipeline
object to
guarantee that the data will be allocated and initialized on the right
memory banks (according to the first touch policy) during the replication
process.
To specify the pinning policy, we defined a syntax to express hwloc
objects
with three different separators:
- Pipeline stage (does not concern
runtime::Sequence
):|
- Replicated stage (= replicated sequence = one thread):
;
- For one thread, the list of pinned
hwloc
objects (= logical or):,
Then, the pinning policy can contains all the available hwloc
objects. Below
is the correspondence between the std::string
and the hwloc
object types:
std::map<std::string, hwloc_obj_type_t> str_to_hwloc_obj =
{
/* global containers */ /* data caches */ /* instruction caches */
{ "GROUP", HWLOC_OBJ_GROUP }, { "L5D", HWLOC_OBJ_L5CACHE }, { "L3I", HWLOC_OBJ_L3ICACHE },
{ "NUMA", HWLOC_OBJ_NUMANODE }, { "L4D", HWLOC_OBJ_L4CACHE }, { "L2I", HWLOC_OBJ_L2ICACHE },
{ "PACKAGE", HWLOC_OBJ_PACKAGE }, { "L3D", HWLOC_OBJ_L3CACHE }, { "L1I", HWLOC_OBJ_L1ICACHE },
{ "L2D", HWLOC_OBJ_L2CACHE }, /* compute units */
{ "L1D", HWLOC_OBJ_L1CACHE }, { "CORE", HWLOC_OBJ_CORE },
{ "PU", HWLOC_OBJ_PU },
};
To specify the index X
of an hwloc
object, the following syntax is used:
OBJECT_X
(ex: PU_5
refers to the logical PU n°5).
Info
CORE
and PU
objects can be confusing. If the CPU cores do not support
SMT, then CORE
and PU
are the same. However, if the CPU cores support
SMT, then the PU
is the hardware thread identifier inside a given CORE
.
Illustrative Examples¶
This section gives some examples to understand how the syntax works. We suppose that we have a CPU with 8 PUs with the same topology as the the Orange Pi 5 Plus SBC presented before.
Example 1¶
Let's suppose we want to setup a 3-stage pipeline with the following characteristics:
- Stage 1 - No replication (= 1 thread):
- Pinned to
PU_0
- Pinned to
- Stage 2 - 4 replications (= 4 threads):
- Thread n°1 is pinned to
PU_4
orPU_5
- Thread n°2 is pinned to
PU_4
orPU_5
- Thread n°3 is pinned to
PU_6
orPU_7
- Thread n°4 is pinned to
PU_6
orPU_7
- Thread n°1 is pinned to
- Stage 3 - No replication (= 1 thread):
- Pinned to
PU_0
,PU_1
,PU_2
orPU_3
- Pinned to
graph LR;
S1T1(Stage 1, thread 1 - pin: PU_0)-->SYNC1;
SYNC1(Sync)-->S2T1;
SYNC1(Sync)-->S2T2;
SYNC1(Sync)-->S2T3;
SYNC1(Sync)-->S2T4;
S2T1(Stage 2, thread 1 - pin: PU_4 or PU_5)-->SYNC2;
S2T2(Stage 2, thread 2 - pin: PU_4 or PU_5)-->SYNC2;
S2T3(Stage 2, thread 3 - pin: PU_6 or PU_7)-->SYNC2;
S2T4(Stage 2, thread 4 - pin: PU_6 or PU_7)-->SYNC2;
SYNC2(Sync)-->S3T1(Stage 3, thread 1 - pin: PU_0, PU_1, PU_2 or PU_3);
In the previous configuration, 6 threads will execute simultaneously (even if the given architecture supports up to 8 executions in parallel).
To instantiate this runtime::Pipeline
, here are the corresponding constructor
parameters:
- Number of replications (= threads) per stage:
{ 1, 4, 1 }
- Enabling pinning per stage:
{ true, true, true }
- Pinning policy:
"PU_0 | PU_4, PU_5; PU_4, PU_5; PU_6, PU_7; PU_6, PU_7 | PU_0, PU_1, PU_2, PU_3"
The previous pinning policy syntax can be compressed a little bit as follow:
- Pinning policy :
"PU_0 | PACKAGE_1; PACKAGE_1; PACKAGE_2; PACKAGE_2 | PACKAGE_0"
Example 2¶
Let's now consider that we want to pin all the threads of the stage 2 on the
PU_4
, PU_5
, PU_6
or PU_7
(this is less restrictive than the previous
example). The pinning strategy for stage 1 and 3 is unchanged.
graph LR;
S1T1(Stage 1, thread 1 - pin: PU_0)-->SYNC1;
SYNC1(Sync)-->S2T1;
SYNC1(Sync)-->S2T2;
SYNC1(Sync)-->S2T3;
SYNC1(Sync)-->S2T4;
S2T1(Stage 2, thread 1 - pin: PU_4, PU_5, PU_6 or PU_7)-->SYNC2;
S2T2(Stage 2, thread 2 - pin: PU_4, PU_5, PU_6 or PU_7)-->SYNC2;
S2T3(Stage 2, thread 3 - pin: PU_4, PU_5, PU_6 or PU_7)-->SYNC2;
S2T4(Stage 2, thread 4 - pin: PU_4, PU_5, PU_6 or PU_7)-->SYNC2;
SYNC2(Sync)-->S3T1(Stage 3, thread 1 - pin: PU_0, PU_1, PU_2 or PU_3);
Here are the corresponding parameters:
- Number of replications (= threads) per stage:
{ 1, 4, 1 }
- Enabling pinning per stage:
{ true, true, true }
- Pinning policy :
"PU_0 | PACKAGE_1, PACKAGE_2 | PACKAGE_0"
With the previous syntax, the 4 threads of the stage 2 will apply the
PACKAGE_1, PACKAGE_2
policy.
Example 3¶
It is also possible to choose the stages we want to pin or not using a vector of
boolean
. Let's suppose we do not want to specify any pinning for the stage 1.
graph LR;
S1T1(Stage 1, thread 1 - no pinning)-->SYNC1;
SYNC1(Sync)-->S2T1;
SYNC1(Sync)-->S2T2;
SYNC1(Sync)-->S2T3;
SYNC1(Sync)-->S2T4;
S2T1(Stage 2, thread 1 - pin: PU_4, PU_5, PU_6 or PU_7)-->SYNC2;
S2T2(Stage 2, thread 2 - pin: PU_4, PU_5, PU_6 or PU_7)-->SYNC2;
S2T3(Stage 2, thread 3 - pin: PU_4, PU_5, PU_6 or PU_7)-->SYNC2;
S2T4(Stage 2, thread 4 - pin: PU_4, PU_5, PU_6 or PU_7)-->SYNC2;
SYNC2(Sync)-->S3T1(Stage 3, thread 1 - pin: PU_0, PU_1, PU_2 or PU_3);
Here are the corresponding parameters:
- Number of replications (= threads) per stage:
{ 1, 4, 1 }
- Enabling pinning per stage:
{false, true, true}
- Pinning policy:
"| PACKAGE_1, PACKAGE_2 | PACKAGE_0"
In this case, the OS will be in charge of pinning the thread of the first stage.
Unpin¶
An unpin function exists and can be called by each thread individually. Once
the unpin
function is triggered the thread will be free to be scheduled by the
OS over all the process units.
Warning
We assume that the user is aware of the computer architecture, uses the
logical indexes of hwloc
and follows the previously defined syntax rules,
otherwise the code will throw exceptions.