# US 2020/0348875 A1 · Method and system for proactive data migration across tiered storage

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2020/0348875 A1
- Document ID: US20200348875A1
- Status: Published application
- Kind: A1
- Application number: 16/403,344
- Filed: 2019-05-03
- Published: 2020-11-05
- Earliest claimed: 2019-05-03
- Assignees: EMC IP Holding Company LLC
- Applicants: EMC IP Holding Company LLC
- Inventors: Jonathan I. Krasner; Jason Jerome Duquette
- Web page: https://patentgrep.com/patent/US20200348875A1

## CPC

- G06F 3/0647: {Migration mechanisms}
- G06F 3/0604: {Improving or facilitating administration, e.g. storage management}
- G06F 3/0649: {Lifecycle management}
- G06F 3/0659: {Command handling arrangements, e.g. command buffers, queues, command scheduling}
- G06F 3/0685: {Hybrid storage combining heterogeneous device types, e.g. hierarchical storage, hybrid arrays}
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- G06N 20/00: Machine learning
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 5/01: Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
- G06N 20/10: using kernel methods, e.g. support vector machines [SVM]
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A method and system for proactive data migration across tiered storage that uses machine learning is shown. The method includes receiving first input-output (TO) event metadata describing a first IO event, and analyzing the first IO event metadata using an optimized learning model (OLM), to obtain a first learning model output. The first learning model output is interpreted to generate a first data migration request, and the first data is proactively migrated from a first storage tier to a second storage tier based on the first data migration request.

## Claims

### Claim 1

A method for proactive data migration across tiered storage, comprising:
receiving first input-output (TO) event metadata describing a first IO event;
analyzing the first IO event metadata using an optimized learning model (OLM), to obtain a first learning model output;
interpreting the first learning model output, to generate a first data migration request; and
migrating, proactively, first data from a first storage tier to a second storage tier based on the first data migration request
aggregating a set of historical IO event metadata describing, respectively, a set of historical IO events;
partitioning the set of historical IO event metadata, to at least obtain a training set comprising a first historical IO event metadata subset;
training a learning model using the training set, to derive optimal learning model parameters; and
adjusting the OLM using at least the optimal learning model parameters, to obtain an adjusted OLM.

### Claim 2 (depends on 1)

The method of claim 1, wherein the IO event metadata comprises observed IO per seconds (TOPS), read percentages, read IO sizes, write IO sizes, and IO response times.

### Claim 3 (depends on 1)

The method of claim 1, wherein the OLM is a recurrent neural network (RNN).

### Claim 4 (depends on 1)

The method of claim 1, wherein the first learning model output comprises a set of data access probabilities for a data set.

### Claim 5 (depends on 1)

The method of claim 1, wherein generating the first data migration request may be performed based on a migration cost metric.

### Claim 6 (depends on 5)

The method of claim 5, wherein the migration cost metric comprises an estimate for time needed to complete migration of the first data from the first storage tier to the second storage tier.

### Claim 7

(canceled)

### Claim 8 (depends on 1)

The method of claim 1, wherein the set of historical IO event metadata comprises the first IO event metadata, wherein the set of historical IO events comprises the first IO event.

### Claim 9 (depends on 1)

The method of claim 1, wherein training the learning model further uses feedback data derived from the first learning model output, to derive the optimal learning model parameters.

### Claim 10 (depends on 1)

The method of claim 1, further comprising:
partitioning the set of historical IO event metadata, to further obtain a validation set comprising a second historical IO event metadata subset;
validating the learning model using the validation set, to derive optimal learning model hyper-parameters; and
adjusting the OLM further using the optimal learning model hyper-parameters, to obtain the adjusted OLM.

### Claim 11 (depends on 10)

The method of claim 10, further comprising:
receiving second IO event metadata describing a second IO event;
analyzing the second IO event metadata using the adjusted OLM, to obtain a second learning model output;
interpreting the second learning model output, to generate a second data migration request; and
migrating, proactively, second data from a third storage tier to a fourth storage tier based on the second data migration request.

### Claim 12 (depends on 1)

The method of claim 1, further comprising:
receiving second IO event metadata describing a second IO event;
analyzing the second IO event metadata using the OLM, to obtain a second learning model output;
interpreting the second learning model output, to generate a second data migration request; and
migrating, proactively, second data from a third storage tier to a fourth storage tier based on the second data migration request.

### Claim 13

A data storage system (DSS), comprising:
an offload domain comprising a first computer processor;
an access prediction service (APS), at least in part, executing on the first computer processor, and programmed to:
receive input-output (TO) event metadata describing an IO event; and
analyze the IO event metadata using an optimized learning model (OLM), to obtain a learning model output;
a central processing unit (CPU) domain operatively connected to the offload domain, and comprising a second computer processor,
wherein the APS further executes, at least in part, on the second computer processor, and is further programmed to:
interpret the learning model output, to generate a data migration request., and
migrate, proactively, data from a first storage tier to a second storage tier based on the data migration request;
aggregating a set of historical IO event metadata describing, respectively, a set of historical IO events;
partitioning the set of historical IO event metadata, to at least obtain a training set comprising a first historical IO event metadata subset;
training a learning model using the training set, to derive optimal learning model parameters; and
adjusting the OLM using at least the optimal learning model parameters, to obtain an adjusted OLM.

### Claim 14 (depends on 13)

The DSS of claim 13, wherein the first computer processor is one selected from a group consisting of a secondary central processing unit (CPU) and a graphics processing unit (GPU).

### Claim 15

(canceled)

### Claim 16 (depends on 13)

The DSS of claim 13, wherein the second computer processor is a primary CPU.

### Claim 17 (depends on 13)

The DSS of claim 13, further comprising:
a physical storage array (PSA) operatively connected to the offload domain and the CPU domain, and comprising a plurality of physical storage devices (PSDs),
wherein the first storage tier comprises a first subset of the plurality of PSDs,
wherein the second storage tier comprises a second subset of the plurality of PSDs.

### Claim 18 (depends on 17)

The DSS of claim 17, wherein the first subset of the plurality of PSDs comprises high-performance data storage, wherein the second subset of the plurality of PSDs comprises low-performance data storage.

### Claim 19 (depends on 17)

The DSS of claim 17, wherein the first subset of the plurality of PSDs comprises low-performance data storage, wherein the second subset of the plurality of PSDs comprises high-performance data storage.

### Claim 20

A non-transitory computer readable medium (CRM) comprising computer readable program code, which when executed by a computer processor, enables the computer processor to:
receive input-output (TO) event metadata describing an IO event;
analyze the IO event metadata using an optimized learning model (OLM), to obtain a learning model output;
interpret the learning model output, to generate a data migration request; and
migrate, proactively, data from a first storage tier to a second storage tier based on the data migration request
aggregating a set of historical IO event metadata describing, respectively, a set of historical IO events;
partitioning the set of historical IO event metadata, to at least obtain a training set comprising a first historical IO event metadata subset;
training a learning model using the training set, to derive optimal learning model parameters; and
adjusting the OLM using at least the optimal learning model parameters, to obtain an adjusted OLM.
