Method for deploying a task in a supercomputer by searching switches for interconnecting nodes
A method for deploying a task includes deploying the task in the supercomputer; executing the task; at the end of the execution of the task, detecting at least one link which is not allocated to any task, and setting each detected link in an inactive state, wherein the link requires a power consumption less than the power consumption required by a link associated with at least one task.
1. A method for deploying a task in a computer cluster forming a supercomputer, the supercomputer comprising
computers, wherein each computer of said computers comprises at least one network interface forming a node; and
a network for interconnecting all of said node from each computer of said computers, comprising
switches, each switch of said switches comprising network interfaces, wherein the switches are arranged in a Parallel Generalized Fat Tree (PGFT) topology having at least four levels between which the switches are distributed, and
links, each link of said links connecting either
said node and said network interfaces of one switch or said switches, or
two network interfaces of two switches respectively;
the method comprising:
deploying the task in the supercomputer, the deploying comprising:
allocating nodes to the task;
scanning the switches in ascending order of level, to find all possible subnetworks, having a same minimum level, interconnecting the nodes that are allocated to said task;
selecting from among the possible subnetworks, a first subnetwork using the most switches already used;
allocating the first subnetwork and the links belonging to said first subnetwork, to the task, each link of the first subnetwork being dedicated to the task; and
implementing inter-node communication routes in the first subnetwork that is allocated;
the method further comprising:
executing the task; and
at an end of the executing of the task:
releasing the links belonging to the first subnetwork allocated to the task that ended;
detecting, among the links that are released, at least one link that is not allocated to any task; and
setting each link of said at least one link that is detected in an idle state, wherein the each link that is in the idle state requires less energy consumption than a link associated with at least one task.
2. The method according to claim 1 , further comprising, for setting said each link in the idle state, switching off the two network interfaces that the link connects such that no current or light flows through the each link.
3. The method according to claim 1 , further comprising, in order to set said each link in the idle state, reducing an operating frequency of at least one of the two network interfaces that said each link connects.
4. The method according to claim 1 , wherein the first subnetwork uses only links which are not allocated to any other already deployed task or which are allocated to fewer than N other tasks already deployed, N being a predefined number equal to one or more.
5. The method according to claim 4 , wherein N is one.
6. A non-transitory computer readable medium comprising a program including instructions for executing steps of a method for deploying a task in a computer cluster forming a supercomputer, when said program is executed on one or more computers,
wherein said supercomputer comprises
said one or more computers, each computer of said one or more computers comprising at least one network interface forming a node;
a network for interconnecting all of said node from said each computer of said one or more computers, comprising
switches, each switch of said switches comprising network interfaces, wherein the switches are arranged in a Parallel Generalized Fat Tree (PGFT) topology having at least four levels between which the switches are distributed, and
links, each link of said links connecting either
said node and said network interfaces of one switch or said switches, or
two network interfaces of two switches respectively;
the method comprising:
deploying the task in the supercomputer, the deploying comprising:
allocating nodes to the task;
scanning the switches in ascending order of level, to find all possible subnetworks, having the same minimum level, interconnecting the nodes that are allocated to said task;
selecting from among the possible subnetworks, a first subnetwork using the most switches already used;
allocating the first subnetwork and the links belonging to said first subnetwork, to the task, each link of the first subnetwork being dedicated to the task; and
implementing inter-node communication routes in the first subnetwork that is allocated;
the method further comprising:
executing the task; and
at an end of the executing of the task:
releasing the links belonging to the first subnetwork allocated to the task that ended;
detecting, among the links that are released, at least one link that is not allocated to any task; and
setting each link of said at least one link that 1s detected in an idle state, wherein the each link that is in the idle state requires less energy consumption than a link associated with at least one task.
7. A computer cluster forming a supercomputer comprising:
computers, each computer of said computers comprising at least one network interface forming a node;
a network for interconnecting all of said node from said each computer of said computers, said network comprising:
switches, each switch of said switches comprising a network interface, wherein the switches are arranged in a Parallel Generalized Fat Tree (PGFT) topology having at least four levels between which the switches are distributed, and
links, each link of said links connecting either
said node and said network interfaces of one switch of said switches, or
two network interfaces of two respective switches of said switches;
the supercomputer further comprising one or more of software and hardware components for deploying a task, configured to implement steps of:
deploying the task in the supercomputer, the deploying comprising:
allocating nodes to the task;
scanning the switches in ascending order of level, to find all possible subnetworks, having the same minimum level, interconnecting the nodes that are allocated to said task;
selecting from among the possible subnetworks, a first subnetwork using the most switches already used;
allocating the first subnetwork and the links belonging to said first subnetwork, to the task, each link of the first subnetwork being dedicated to the task; and
implementing inter-node communication routes in the first subnetwork that is allocated;
executing the task; and
at an end of said executing of the task:
releasing the links belonging to the first subnetwork allocated to the task that ended;
detecting, among the links that are released, at least one link that is not allocated to any task, and
setting each link of said at least one link that is detected in an idle state, wherein the each link that is in the idle state requires less energy consumption than a link associated with at least one task.