Increasing performance of write throughput using machine learning

Patent metadata

Field	Value
Publication number	US-10853246-B2
Application number	US-201816175179-A
Country	US
Kind code	B2
Filing date	Oct 30, 2018
Priority date	Oct 30, 2018
Publication date	Dec 1, 2020
Grant date	Dec 1, 2020

Abstract

Techniques for processing data may include: determining a first amount denoting an amount of write pending data stored in cache to be redirected through storage class memory (SCM) when destaging cached write pending data from the cache; performing first processing that destages write pending data from the cache, the first processing including: selecting, in accordance with the first amount, a first portion of write pending data that is destaged from the cache and stored in the SCM and a second portion of write pending data that is destaged directly from the cache and stored on one or more physical storage devices providing back-end non-volatile physical storage; and subsequent to storing the first portion of write pending data to the SCM, transferring the first portion of write pending data from the SCM to the one or more physical storage devices providing back-end non-volatile physical storage.

First claim

What is claimed is:

1. A method of processing data comprising:
determining a first amount denoting an amount of write pending data stored in cache to be redirected through storage class memory (SCM) when destaging cached write pending data from the cache;
performing first processing that destages write pending data from the cache, the first processing including:
selecting, in accordance with the first amount, a first portion of write pending data that is destaged from the cache and stored in the SCM and a second portion of write pending data that is destaged directly from the cache and stored on one or more physical storage devices providing back-end non-volatile physical storage; and
subsequent to storing the first portion of write pending data to the SCM, transferring the first portion of write pending data from the SCM to the one or more physical storage devices providing back-end non-volatile physical storage;

determining, using a Markov chain state machine (MCSM) included in a predictor, a predicted probabilistic risk of transitioning from a current state to a high state of heavy write throughput; and
providing the probabilistic risk from the predictor as an input to a controller, wherein the MCSM uses a matrix of state transition probabilities to determine the predicted probabilistic risk of transitioning from the current state to the high state of heavy write throughput.

2. The method of claim 1 , wherein a relative performance ranking ranks the cache as having a higher performance than the SCM, and the SCM as having a higher performance than the one or more physical storage devices providing back-end non-volatile physical storage.

3. The method of claim 2 , wherein the SCM is byte addressable.

4. The method of claim 1 , wherein the controller determines the first amount in accordance with a plurality of inputs including the predicted probabilistic risk of transitioning from the current state to the high state of heavy write throughput.

5. The method of claim 4 , wherein the first plurality of inputs to the controller further include a set point denoting a current acceptable level of probabilistic risk of transitioning to the high state of heavy write throughput, and a delta risk value that is a difference between the current acceptable level of probabilistic risk and the predicted probabilistic risk of transitioning from the current state to the high state of heavy write throughput.

6. The method of claim 5 , wherein the first amount is expressed as a first percentage denoting a percentage of write pending data to be destaged indirectly through the SCM.

7. The method of claim 5 , wherein the controller determines the first amount for a current time period in accordance with the delta risk value for the current time period and a first weighting constant, a sum of delta risk values from time periods prior to the current time period and a second weighting constant, and a rate of change in delta risk values for the current time period and an immediately prior time period and a third weighting constant.

8. The method of claim 1 , wherein the matrix of state transition probabilities used by the MCSM is determined by observed transitions between states in accordance with observed write throughput.

9. The method of claim 8 , wherein the MCSM includes three states and the matrix includes nine entries denoting transitions from a current state denoting one of the three states to a next state denoting one of the three states.

10. The method of claim 9 , wherein each of the three states is associated with a different range of write throughput.

11. A system comprising:
a processor; and
a memory comprising code stored thereon that, when executed, performs a method of processing data comprising:
determining a first amount denoting an amount of write pending data stored in cache to be redirected through storage class memory (SCM) when destaging cached write pending data from the cache;
performing first processing that destages write pending data from the cache, the first processing including:
selecting, in accordance with the first amount, a first portion of write pending data that is destaged from the cache and stored in the SCM and a second portion of write pending data that is destaged directly from the cache and stored on one or more physical storage devices providing back-end non-volatile physical storage; and
subsequent to storing the first portion of write pending data to the SCM, transferring the first portion of write pending data from the SCM to the one or more physical storage devices providing back-end non-volatile physical storage;

determining, using a Markov chain state machine (MCSM) included in a predictor, a predicted probabilistic risk of transitioning from a current state to a high state of heavy write throughput; and
providing the probabilistic risk from the predictor as an input to a controller, wherein the MCSM uses a matrix of state transition probabilities to determine the predicted probabilistic risk of transitioning from the current state to the high state of heavy write throughput.

12. A non-transitory computer readable medium comprising code stored thereon that, when executed, performs a method of processing data comprising:
determining a first amount denoting an amount of write pending data stored in cache to be redirected through storage class memory (SCM) when destaging cached write pending data from the cache;
performing first processing that destages write pending data from the cache, the first processing including:
selecting, in accordance with the first amount, a first portion of write pending data that is destaged from the cache and stored in the SCM and a second portion of write pending data that is destaged directly from the cache and stored on one or more physical storage devices providing back-end non-volatile physical storage; and
subsequent to storing the first portion of write pending data to the SCM, transferring the first portion of write pending data from the SCM to the one or more physical storage devices providing back-end non-volatile physical storage;

determining, using a Markov chain state machine (MCSM) included in a predictor, a predicted probabilistic risk of transitioning from a current state to a high state of heavy write throughput; and
providing the probabilistic risk from the predictor as an input to a controller, wherein the MCSM uses a matrix of state transition probabilities to determine the predicted probabilistic risk of transitioning from the current state to the high state of heavy write throughput.

13. The non-transitory computer readable medium of claim 12 , wherein a relative performance ranking ranks the cache as having a higher performance than the SCM, and the SCM as having a higher performance than the one or more physical storage devices providing back-end non-volatile physical storage.

14. The non-transitory computer readable medium of claim 13 , wherein the SCM is byte addressable.

15. The non-transitory computer readable medium of claim 12 , wherein the controller determines the first amount in accordance with a plurality of inputs including the predicted probabilistic risk of transitioning from the current state to the high state of heavy write throughput.

16. The non-transitory computer readable medium of claim 15 , wherein the first plurality of inputs to the controller further include a set point denoting a current acceptable level of probabilistic risk of transitioning to the high state of heavy write throughput, and a delta risk value that is a difference between the current acceptable level of probabilistic

Assignees

• Emc Ip Holding Co Llc

Inventors

• Randolph Benjamin A
• Martin Owen

Classifications

• G06N7/01

  Probabilistic graphical models, e.g. probabilistic networks · CPC title

• G06F2212/502

  using adaptive policy · CPC title

• G06F12/0897

  with two or more cache hierarchy levels (with multilevel cache hierarchies G06F12/0811) · CPC title

• G06F3/0671

  In-line storage system · CPC title

• G06F12/0888

  using selective caching, e.g. bypass · CPC title