Load balancing technique for a storage array
A technique for load balancing uses heuristic-based algorithms with respect to input/output (I/O) latency of workloads destined to storage devices, e.g., solid state drives (SSDs), of a storage array attached to a storage system. Illustratively, “front-end” requests received from a host result in a back-end workload as those requests are processed by a storage I/O stack of the storage system and stored on the storage array. Accordingly, the technique maintains a consistent latency for the host requests (front-end) to control latency for the back-end workload. The load balancing technique illustratively load balances fixed (back-end) workloads having similar I/O sizes and I/O patterns. Illustratively, the technique balances the workloads across a plurality of storage ports over one or more I/O paths to the SSDs. Access to the SSDs may then be distributed among the storage ports.
1. A method comprising:
receiving from a host, a plurality of write requests directed towards a logical unit (LUN), each write request having data, the write request processed at a storage system having a memory coupled to a processor via a memory bus, the processor coupled to at least one storage controller attached to at least one shelf of solid state drives (SSDs);
generating an input/output (I/O) workload from the plurality of write requests, the I/O workload including a de-duplication operation;
looking ahead into an I/O queue having the I/O workload to predict an expected load on each I/O channel of a plurality of I/O channels included in the at least one storage controller; and
controlling a back-end latency of the I/O workload to the SSDs by load balancing the I/O workload among the plurality of I/O channels based on the expected work load on each I/O channel, thereby for the plurality of write requests maintaining a predictable front-end latency to the host.
2. The method of claim 1 wherein predicting an expected load on each I/O channel of a plurality of I/O channels further comprises:
using a read/write workload ratio, wherein a first latency associated with a read operation and a second latency associated with a write operation are known, and wherein the I/O workload includes the read operation and the write operation.
3. The method of claim 1 , wherein each SSD is associated with a primary I/O path and a secondary I/O path to the at least one storage controller, and wherein the load balancing of the I/O workload includes sending data across the primary I/O path and the secondary I/O path for each SSD.
4. The method of claim 1 wherein the I/O workload includes an unexpected read operation in response to the de-duplication operation.
5. The method of claim 1 , wherein each SSD is associated with a primary I/O path and a secondary I/O path to the at least one storage controller, wherein the load balancing of the I/O workload includes sending data across the primary I/O path for each SSD, and wherein the secondary I/O path is a failover I/O path for each SSD.
6. The method of claim 1 further comprising:
reading a temperature of each respective SSD; and
in response to the temperature of the respective SSD being outside of an operating temperature range, issuing a warmup I/O request to the respective SSD.
7. The method of claim 1 wherein load balancing the I/O workload includes calculating a number of I/O operations per second of the workload for each I/O channel.
8. The method of claim 1 wherein load balancing the I/O workload includes calculating an I/O latency histogram of the I/O workload for a plurality of durations.
9. The method of claim 1 wherein load balancing the I/O workload includes examining a reads pending count and a writes pending count indicating a respective number of read and write requests associated with the I/O queue.
10. The method of claim 5 further comprising:
monitoring stream-open durations for a first SSD and a second SSD, wherein the first and second SSDs are multi-streaming; and
in response to the first SSD having stream-open durations greater than an average stream-open duration and the second SSD having stream-open durations less than the average stream-open duration, exchanging a first primary I/O path associated with the first SSD with a second primary I/O path associated with the second SSD.
11. A method comprising:
receiving from a host, a plurality of write request directed towards a logical unit (LUN), each write request having data, the write request processed at a storage system having a memory coupled to a processor via a memory bus, the processor coupled to at least one storage controller attached to at least one shelf of multi-stream solid state drives (SSDs);
generating an input/output (I/O) workload from the plurality of write requests, the workload including the data and metadata associated with the data; and
controlling a back-end latency of the I/O workload to the SSDs by load balancing the I/O workload among a plurality of I/O channels included in the at least one storage controller such that the plurality of write requests maintain a predictable front-end latency to the host, wherein a first stream for each SSD is used for the data and a second stream different from the first stream for each SSD is used for the metadata.
12. A system comprising:
a storage system having a memory connected to a processor via a bus;
at least one storage shelf coupled to the processor via at least one storage controller, each storage shelf having a plurality of solid state drives (SSDs);
a storage I/O stack executing on the processor of the storage system, the storage I/O stack configured to:
receive from a host, a plurality of write requests directed towards a logical unit (LUN), each write request having data;
generate an input/output (I/O) workload from the plurality of write requests, the I/O workload including a de-duplication operation;
look ahead into an I/O queue having the I/O workload to predict an expected load on each I/O channel of a plurality of I/O channels included in the at least one storage controller; and
control a back-end latency of the I/O workload to the SSDs by load balancing the I/O workload among the plurality of I/O channels based on the expected work load on each I/O channel, thereby for the plurality of write requests maintaining a predictable front-end latency to the host.
13. The system of claim 12 wherein the storage I/O stack configured to predict an expected load on each I/O channel of a plurality of I/O channels is further configured to:
use a read/write workload ratio, wherein a first latency associated with a read operation and a second latency associated with a write operation are known, and wherein the I/O workload includes the read operation and the write operation.
14. The system of claim 12 , wherein each SSD is associated with a primary I/O path and a secondary I/O path to the at least one storage controller, and wherein the load balancing of the I/O workload includes sending data across the primary I/O path and the secondary I/O path for each SSD.
15. The system of claim 12 the I/O workload includes an unexpected read operation in response to the de-duplication operation.
16. The system of claim 12 , wherein each SSD is associated with a primary I/O path and a secondary I/O path to the at least one storage controller, and wherein the load balancing of the I/O workload includes sending data across the primary I/O path for each SSD, and wherein the secondary I/O path is a failover I/O path for each SSD.
17. The system of claim 12 wherein the storage I/O stack is further configured to:
read a temperature of each respective SSD; and
in response to the temperature of the respective SSD being outside of an operating temperature range, issue a warmup I/O request to the respective SSD.
18. The system of claim 12 wherein load balancing the I/O workload includes calculating a number of I/O operations per second of the workload for each I/O channel.
19. The system of claim 12 wherein load balancing the I/O workload includes calculating a I/O latency histogram of the I/O workload for a plurality of durations.
20. The system of claim 12 wherein the load balancing the I/O workload includes examining a reads pending counter and a writes pending counter indicating a respective number of read and write requests associated with the I/O queue.