CPU overcommit with guest idle polling
A system for CPU overcommit with guest idle-polling includes a memory, a plurality of host CPUs in communication with the memory, a virtual machine, and a hypervisor. The plurality of host CPUs include a first host CPU and a second host CPU. The virtual machine includes a virtual central processing unit (VCPU) and a guest. In an example, the VCPU executes on the first host CPU. The guest records in a location of guest memory a value indicating that the VCPU is in an idle-polling state. Then, when the hypervisor detects that a new task needs to be executed, the hypervisor reads the value in the location of the guest memory and determines that the VCPU is in the idle-polling state. Then, the hypervisor schedules the new task to the first host CPU.
1. A method comprising:
sending, by a first guest to a hypervisor executing on a plurality of host central processing units (CPUs), an address of a location in guest memory and information specifying which value indicates that a given virtual central processing unit (VCPU) is in an idle-polling state during a boot of the first guest,
recording, by the hypervisor, the address of the location and the information,
executing, by the hypervisor, a first VCPU of a first virtual machine on a first host CPU, wherein the plurality of host CPUs include the first host CPU and a second host CPU;
determining, by the hypervisor, that the first VCPU is in the idle-polling state by reading a value in the location and comparing the value in the location with the information; and
scheduling, by the hypervisor, a new task to the first host CPU or the second host CPU.
2. The method of claim 1 , wherein the first guest in the first virtual machine records in the location the value indicating that the first VCPU is in the idle-polling state.
3. The method of claim 2 , scheduling, by the hypervisor, the new task to the first host CPU.
4. The method of claim 3 , further comprising responsive to determining that the first VCPU is in the idle-polling state, halting, by the hypervisor, the first VCPU before scheduling the new task to the first host CPU.
5. The method of claim 4 , wherein the first guest modifies the value in the location when the first VCPU starts executing a second task after the first VCPU is halted by the hypervisor, wherein the modified value indicates that the first VCPU is in a non-idle-polling state.
6. The method of claim 1 , wherein the new task comprises a second VCPU.
7. The method of claim 1 , wherein the new task comprises a write/read instruction or a networking related task.
8. The method of claim 1 , further comprising sending, by the hypervisor, a request to the first guest that the first guest change the value in the location whenever a state of the first VCPU changes between the idle-polling state and a non-idle-polling state.
9. The method of claim 1 , further comprising receiving, by the hypervisor from the first guest, a notification that the first guest records a value in the location whenever a state of the first VCPU changes between the idle-polling state and a non-idle-polling state.
10. The method of claim 1 , further comprising recording, by the hypervisor, the address of the location and the information in hypervisor memory.
11. The method of claim 10 , wherein the hypervisor memory includes a task structure, wherein the hypervisor records the address of the location and the information in the task structure.
12. A system comprising:
a plurality of host central processing units (CPUs), wherein the plurality of host CPUs include a first host CPU and a second host CPU;
a first guest configured to send to a hypervisor executing on the plurality of host CPUs, an address of a location in guest memory and information specifying which value indicates that a given virtual central processing unit (VCPU) is in an idle-polling state during a boot of the first guest; and
the hypervisor configured to:
record the address of the location and the information,
execute a first VCPU of a first virtual machine on the first host CPU;
determine that the first VCPU is in the idle-polling state by reading a value in the location and comparing the value in the location with the information; and
schedule a new task to the first host CPU or the second host CPU.
13. The system of claim 12 , wherein the new task comprises a second VCPU.
14. The system of claim 12 , wherein the hypervisor is further configured to send a request to the first guest that the first guest change the value in the location whenever a state of the first VCPU changes between the idle-polling state and a non-idle-polling state.
15. The system of claim 12 , wherein the hypervisor is further configured to record the address of the location and the information in a task structure in hypervisor memory.
16. A non-transitory machine readable medium storing instructions, which when executed by at least one of a plurality of host central processing units (CPUs), cause the plurality of host CPUs to:
send, by a first guest to a hypervisor executing on the plurality of host CPUs, an address of a location in guest memory and information specifying which value indicates that a given virtual central processing unit (VCPU) is in an idle-polling state during a boot of the first guest,
record, by the hypervisor executing on the plurality of host CPUs, the address of the location and the information,
execute, by the hypervisor, a first VCPU of a first virtual machine on a first host CPU, wherein the plurality of host CPUs include the first host CPU and a second host CPU;
determine, by the hypervisor, that the first VCPU is in the idle-polling state by reading a value in the location and comparing the value in the location with the information; and
schedule, by the hypervisor, a new task to the first host CPU or the second host CPU.