Admission control for latency-critical remote procedure calls in datacenters
A distributed sender driven Admission Control System (ACS) is described herein, leveraging Weighted-Fair Quality of Service (QoS) queues, found in standard NICs and switches, to guarantee RPC level latency service level objectives (SLOs) by a judicious selection of QoS weights and traffic-mix across QoS queues. ACS installs cluster-wide RPC latency SLOs by mapping LS RPCs to higher weight QoS queues, and coping with overloads by adaptively apportioning LS RPCs amongst QoS queues based on measured completion times for each queue. When the network demand spikes unexpectedly to predetermined threshold percentage of provisioned capacity, ACS achieves a latency SLO that is significantly lower than the state-of-art congestion control at the 99.9th-p and admits significantly more RPCs meeting SLO target when RPC sizes are not aligned with priorities.
1. A method, comprising:
mapping, by one or more processors of a host device, priority class of a remote procedure call (RPC) to a network quality of service (QoS) level;
receiving, from an application, one or more RPCs; and
controlling, by the one or more processors of the host device, at RPC granularity, admission of a given RPC of the one or more RPCs to a particular QoS level using an admit probability for the particular QoS level, wherein the controlling comprises downgrading the given RPC from a requested QoS level to a lower QoS level; and
adjusting the admit probability for the requested QoS level based on measuring network latency of RPCs processed at the requested QoS level.
2. The method of claim 1 , wherein the priority class is one of a plurality of possible priority classes, and wherein the QoS level is one of a plurality of possible QoS levels.
3. The method of claim 1 , wherein each of the one or more RPCs received from the application includes an indication of priority class.
4. The method of claim 3 , wherein the indication of priority class is requested by the application.
5. The method of claim 3 , wherein controlling admission of the given RPC comprises determining, by the one or more processors, whether to admit the given RPC on a QoS level that is mapped to the indicated priority class.
6. The method of claim 1 , wherein controlling admission of the given RPC comprises:
issuing the given RPC at the lower QoS level.
7. The method of claim 1 , wherein downgrading the given RPC comprises automatically downgrading non-admitted RPCs to a lowest QoS level.
8. The method of claim 1 , further comprising communicating downgrade information to an application from which the given RPC was received.
9. The method of claim 1 , wherein controlling admission of the given RPC comprises:
identifying service level objectives for the network latency;
determining, based on the identified service level objectives and the measured network latency, whether to admit the given RPC to the particular QoS level.
10. The method of claim 9 , wherein the adjusting, by the one or more processors, comprises adjusting the admit probability per destination-host.
11. The method of claim 1 , wherein controlling admission of the given RPC comprises:
measuring RPC latency for each QoS level; and
when latency for priority class RPCs is at or above a predetermined threshold, adaptively
setting QoS code points on network traffic.
12. The method of claim 11 , wherein adaptively setting QoS code points comprises:
identifying a service level objective (SLO); and
downgrading out-of-profile traffic to the lower QoS class such that traffic admitted in higher
QOS classes meets the identified SLO.
13. The method of claim 1 , further comprising enforcing, by one or more switches, QoS using weighted fair queues.
14. The method of claim 1 , wherein the mapping comprises modifying packet headers.
15. A system, comprising:
memory; and
one or more processors in communication with the memory, the one or more processors configured to:
map priority class of a remote procedure call (RPC) to a network quality of service (QoS) level;
receive, from an application, one or more RPCs; and
control, at RPC granularity, admission of a given RPC of the one or more RPCs to a particular QoS level using an admit probability for the particular QoS level, wherein the controlling comprises downgrading the given RPC from a requested QoS level to a lower QoS level; and
adjust the admit probability for the requested QoS level based on measuring network latency of RPCs processed at the requested QoS level.
16. The system of claim 15 , wherein the priority class is one of a plurality of possible priority classes, and wherein the QoS level is one of a plurality of possible QoS levels.
17. The system of claim 15 ,
wherein each of the one or more RPCs received from the application includes an indication of priority class requested by the application, and
wherein controlling admission of the given RPC comprises determining, by the one or more processors, whether to admit the given RPC on a QoS level that is mapped to the indicated priority class.
18. The system of claim 15 , wherein in controlling admission of the given RPC, the one or more processors are further configured to:
issue the given RPC at the lower QoS level.
19. The system of claim 15 , wherein controlling admission of the given RPC comprises:
identifying service level objectives for the network latency;
determining, based on the identified service level objectives and the measured network latency, whether to admit the given RPC to the particular QoS level.
20. A non-transitory computer-readable medium storing instructions executable by one or more processors for performing a method, comprising:
mapping priority class of a remote procedure call (RPC) to a network quality of service (QOS) level;
receiving one or more RPCs; and
controlling, at RPC granularity, admission of a given RPC of the one or more RPCs to a particular QoS level using an admit probability for the particular QoS level, wherein the controlling comprises downgrading the given RPC from a requested QoS level to a lower QoS level; and
adjusting the admit probability for the requested QoS level based on measuring network latency of RPCs processed at the requested QoS level.