“Accessibility is not a bolt on. It’s something that must be built in to every product we make so that our products work for everyone. Only then will we empower every person and every organization on the planet to achieve more. This is the inclusive culture we aspire to create” – Satya Nadella 

Our journey in accessible technology is grounded in a shared conviction at Microsoft. As product makers, we believe in the obligation to build technology that truly empowers people, fostering an inherently equitable experience. 

In this pursuit, we’ve embraced a mindset we call “shift left,” incorporating accessibility at every stage and right from the inception of designing and building our products.  

Reflecting on my tenure, I’ve had the privilege of contributing to some of the world’s most impactful technologies, particularly with Office and Windows. Traditionally, these products were well established with years of development and only later received our focused attention on accessibility.  

However, Copilot presented a rare opportunity for us to incorporate accessibility right from the inception of the product and therefore “shift left”, the entire design and development process. And what’s more, AI technologies like Copilot brought a unique opportunity to reshape how humans interact with computers in a way that makes the experience MORE equitable and transformative for all. 

Now, integrated into our Microsoft 365 Core Apps, Copilot brings forth exciting capabilities. Our goal is to bridge the gap between your interaction with technology and how you express your ideas, making the user experience more inclusive and empowering for all. 

At Copilot’s core lies a commitment to equity, underscoring our ongoing dedication to fostering a technology landscape that truly serves every individual, ensuring that no one is left behind. 

Equity at Copilot’s Core 

We are actively shifting left in our product development by making accessibility a core part of Copilot’s design and functionality. Copilot is designed to work well with assistive technologies, such as screen readers, magnifiers, contrast themes, and voice input, and to provide a seamless and intuitive user experience. 

But in addition to that, Copilot is a tool designed to be accessible itself. 

In this process, we have collaborated and co-innovated with a diverse set of customers, 600+ members of Microsoft’s employee disability community, partners in research and design, and the commitment of engineers and product makers to listen and be accountable.  

To illustrate how Copilot can enhance accessibility, I want to share with you some highlights from engaging with participants who had early access to Copilot: 

Drafting emails: Copilot can help create 1st drafts in a matter of minutes. This can be especially helpful to those who have more limited mobility and challenges typing. You can generate different versions of the email with different levels of formality and detail and ask Copilot to check their calendar and suggest a meeting time, just with a few clicks of a button.  
Using voice commands: With Copilot, you can create entire PowerPoint presentations just using your voice. Just tell Copilot about what you want to create, and Copilot can generate relevant graphics and notes for slides. 

These examples demonstrate how Copilot can save users time and effort, as well as help them express their ideas and communicate their expertise more effectively. They also show how Copilot can adapt to their preferences and needs and provide them with a supportive partner that can enhance their communication and productivity. 

In addition to some of these areas of feedback, we also conducted a deep dive study with those of the neurodiverse community. The neurodiverse community (which makes up 10-20% of the world) faces common challenges that we can all relate to, but their lives are disproportionately affected by them. Examples include planning, focus, procrastination, communication, reading ease and comprehension, being overly literal, and fatigue. 

For the neurodiverse community, our study showed that Copilot can be a powerful ally, offering assistance in overcoming these challenges. It serves as a facilitator for thought organization, acts as a springboard for writing tasks, aids in surmounting task initiation barriers, and assists in processing extensive information found in documents, messages, or emails. 

Members of the community reported Copilot helping their communication effectiveness by distilling action items from team meetings and documents, generating  summaries, adjusting the tone and context of their content, and bridging communication gaps. 

As one of the participants in the study said, “For me, Copilot itself is accessibility. Having Copilot is like putting on glasses after I’ve been squinting my entire career. It is equity and I think as a neurodivergent individual, I can’t imagine going back.” 

Making Accessible Content with Ease 

On our journey to create products that are truly inclusive, we’re also empowering document authors to shift left and build better authoring habits by catching accessibility errors early in the doc creation process. Ensuring that your content is comprehensible to all individuals, irrespective of their visual abilities or preferences, is a crucial component of accessibility. To assist you in this endeavor, we have created the Accessibility Assistant, a robust tool that can detect and resolve potential problems in your documents, emails, and presentations. You can access the Accessibility Assistant from the Review tab in Word, Outlook, and PowerPoint. 

New Features of the Accessibility Assistant include some of the following highlights: 

The in-canvas notifications for readability is a feature that notifies you of accessibility hurdles for common issues, such as text color not meeting the Web Content Accessibility Guidelines (WCAG) color contrast ratio or images lacking descriptions. You can use the inclusive color picker to choose an appropriate color from the suggested options and utilize the automatically generated image descriptions to provide alt-text, making it easier to create accessible content. 
Quick fix card for multiple issues: This feature allows you to fix several issues of the same type with fewer clicks. For example, you can change the color of all the text that has low contrast in your document. 
Per-slide toggle for PowerPoint: This feature enables you to view and fix the accessibility issues for each slide individually, instead of seeing them by categories. This can help you focus on your own slides and collaborate with others more easily. 

These capabilities are designed to help you create accessible content faster and easier, and to ensure that everyone can access and enjoy your work. The Accessibility Assistant for Word Desktop has started rolling out to Insider Beta Channel users running Version 2012 (Build 17425.2000) or later. This feature for Outlook Desktop will be available in Insider Beta Channel by April 2024, followed by release to PowerPoint Desktop this summer 

Our Commitment 

At Microsoft, we believe that everyone has something valuable to offer, and that diversity of perspectives and experiences can enrich our products and services. That’s why we are committed to empowering everyone to achieve more, fostering an inherently equitable experience. Copilot is one of the ways that we are fulfilling this commitment, by providing a supportive partner that can help you with common challenges, enhance your communication, and bridge the gap between your interaction with technology and how you express your ideas. 

But we also know that we are not done yet. We are still on a journey of understanding how AI and LLMs will continue to evolve and make the world a more equitable place. We are constantly learning from our customers, partners, and the disability community, and we are always looking for ways to improve our accessibility features and functionality. We welcome your feedback and suggestions on how we can make Copilot better for you and for everyone. 

To learn more about Copilot and how to get started, please visit the Copilot website or the Copilot support page. To learn more about accessibility at Microsoft and how to access our accessibility features, please visit the Microsoft Accessibility website or the Disability Answer Desk. And to share your feedback or suggestions on Copilot, please use the feedback button (thumbs up or down). 

Together, we can make the world a more equitable place for everyone. 

​Microsoft Tech Community – Latest Blogs –Read More 

Author introduction

Hi, I am Saira Shaik, Working Principal customer success account manager at Microsoft India.

This article will provide guidance to the customers who wants to Optimize their Azure costs by providing tools and resources to help customers to save cost, Understand and forecast your costs, Cost optimize workloads and Control costs.

Explore tools and resources to help you save

Find out about the tools, offers, and guidance designed to help you manage and optimize your Azure costs. Learn how to understand and forecast your bill, optimize workload costs, and control your spending.

8 ways to optimize the cost

1.     Shut down unused resources.

Identify idle virtual machines (VMs), ExpressRoute circuits, and other resources with Azure Advisor. Get recommendations on which resources to shut down and see how much you would save.

Useful Links

Reduce service costs using Azure Advisor – Azure Advisor | Microsoft Learn

2.     Right-size underused resources

Find underutilized resources with Azure Advisor—and get recommendations on how to reduce your spend by reconfiguring or consolidating them.

Useful Links

Reduce service costs using Azure Advisor – Azure Advisor | Microsoft Learn

3.     Add an Azure savings plan for compute for dynamic workloads

Save up to 65 percent off pay-as-you-go pricing when you commit to spend a fixed hourly amount on compute services for one or three years.

Useful Links

Azure Savings Plan Savings – youtube.com/playlist?list=PLlrxD0HtieHjd-zn7u09YoGJY18ZrN1Hq

Introduction to Azure savings plan for compute (youtube.com)
Understanding your Azure savings plan recommendations (youtube.com)
How Azure savings plan is applied to a customer’s compute environment (youtube.com)
Azure Savings Plan for Compute | Microsoft Azure

4.     Reserve instances for consistent workloads

Get a discount of up to 72 percent over pay-as-you-go pricing on Azure services when you prepay for a one- or three-year term with reservation pricing.Get a discount of up to 72 percent over pay-as-you-go pricing on Azure services when you prepay for a one- or three-year term with reservation pricing.

Useful Links

Reservations | Microsoft Azure
Advisor Clinic: Lower costs with Azure Virtual Machine reservations (youtube.com)
Model virtual machine costs with the Azure Cost Estimator Power BI Template (youtube.com)

5.     Take advantage of the Azure Hybrid Benefit

AWS is up to five times more expensive than Azure for Windows Server and SQL Server. Save when you migrate your on-premises workloads to Azure.

Useful Links

Azure Hybrid Benefit—hybrid cloud | Microsoft Azure
Reduce costs and increase SQL license utilization using Azure Hybrid Benefit (youtube.com)
Managing and Optimizing Your Azure Hybrid Benefit Usage (With Tools!) – Microsoft Community Hub

6.     Configure autoscaling

Save by dynamically allocating and de-allocating resources to match your performance needs.

Useful Links

Autoscaling guidance – Best practices for cloud applications | Microsoft Learn

7.     Choose the right Azure compute service

Azure offers many ways to host your code. Operate more cost efficiently by selecting the right compute service for your application.

Useful Links

Choose an Azure compute service – Azure Architecture Center | Microsoft Learn
Armchair Architects: Exploring the relationship between Cost and Architecture (youtube.com)

8.     Set up budgets and allocate costs to teams and projects

Create and manage budgets for the Azure services you use or subscribe to—and monitor your organization’s cloud spending—with Microsoft Cost Management.

Useful Links

Tutorial – Create and manage budgets – Microsoft Cost Management | Microsoft Learn
The Cloud Clinic: Use tagging and cost management tools to keep your org accountable (youtube.com)

Understand and forecast your costs

Monitor and analyze your Azure bill with Microsoft Cost Management. Set budgets and allocate spending to your teams and projects.
Estimate the costs for your next Azure projects using the Azure pricing calculator and the Total Cost of Ownership (TCO) calculator.
Successfully build your cloud business case with key financial and technical guidance from Azure.

Useful Links

FinOps toolkit – Kick start your FinOps efforts (microsoft.github.io)
Azure Savings Dashboard – Microsoft Community Hub
Azure Cost Management Dashboard – Microsoft Community Hub

Cost optimize your workloads

Follow your Azure Advisor best practice recommendations for cost savings.

Review your workload architecture for cost optimization using the Microsoft Azure Well-Architected Review assessment and the Microsoft Azure Well-Architected Framework design documentation, well architected cost optimization implementation – Customer Offerings: Well-Architected Cost Optimization Implementation – Microsoft Community Hub

Save with Azure offers and licensing terms such as the Azure Hybrid Benefit, paying in advance for predictable workloads with reservations, Azure Spot Virtual Machines, Azure savings plan for compute, and Azure dev/test pricing.

Control your costs

Mitigate cloud spending risks by implementing cost management governance best practices at your company using the Microsoft Cloud Adoption Framework for Azure.

Implement cost controls and guardrails for your environment with Azure Policy.

​Microsoft Tech Community – Latest Blogs –Read More 

Want to learn more about Generative AI and Microsoft Copilot?  

Microsoft is launching a Learn Live Series called “Getting Started with Microsoft Copilot for Security.” This weekly online seminar series will run from March 19th through April 9th and will review skill development resources and discuss topics related to AI and Copilot for Security. 

Hosts Edward Walton, Andrea Fisher, and Rod Trent will guide you through four topics each with a corresponding Microsoft Learn module designed to help anyone interested in getting users ready for Microsoft Copilot for Security. 

Fundamentals of Generative AI​ 

March 19th 12:00 pm – 1:30 pm PDT 

In this session, you will explore the way in which large language models (LLMs) enable AI applications and services to generate original content based on natural language input. You will also learn how generative AI enables the creation of AI-powered copilots that can assist humans in creative tasks. In this episode, you will: 

Learn about the kinds of solutions AI can make possible and considerations for responsible AI practices 
Understand generative AI’s place in the development of artificial intelligence 
Understand large language models and their role in intelligent applications 
Describe how Azure OpenAI supports intelligent application creation 
Describe examples of copilots and good prompts 

Fundamentals of Responsible Generative AI 

March 27th 12:00 pm –1:30 pm PDT 

Generative AI enables amazing creative solutions but must be implemented responsibly to minimize the risk of harmful content generation. In this episode, you will: 

Describe an overall process for responsible generative AI solution development 
Identify and prioritize potential harms relevant to a generative AI solution 
Measure the presence of harms in a generative AI solution 
Mitigate harms in a generative AI solution 
Prepare to deploy and operate a generative AI solution responsibly 

Get started with Microsoft Security Copilot​ 

April 2nd 12:00pm – 1:30 pm PDT 

Get acquainted with Microsoft Copilot for Security. You will be introduced to some basic terminology, how Microsoft Copilot for Security processes prompts, the elements of an effective prompt, and how to enable the solution. In this episode, you will: 

Describe what Microsoft Copilot for Security is. 
Describe the terminology of Microsoft Copilot for Security. 
Describe how Microsoft Copilot for Security processes prompt requests. 
Describe the elements of an effective prompt 
Describe how to enable your Microsoft Copilot for Security solution. 

Describe the core features of Microsoft Security Copilot 

April 9th 12:00 pm – 1:30 pm PDT  

Microsoft Copilot for Security has a rich set of features. Learn about available plugins that enable integration with various data sources, promptbooks, the ways you can export and share information from Copilot for Security, and much more. In this episode, you will: 

Describe the features available in the standalone experience. 
Describe the services to which Copilot for Security can integrate. 
Describe the embedded experience 

Jump-start your Copilot for Security journey and join us for the Learn Live series starting on Tuesday, March 19th!  

​Microsoft Tech Community – Latest Blogs –Read More 

Guidance for using mysqlslap to simulate client load and measure performance

Introduction

Mysqlslap is a diagnostic program included with the MySQL server binary that you can use to emulate client load for a MySQL server and report the timing of each stage. Mysqlslap works as if multiple clients are accessing the server simultaneously.

In this post, I’ll show you how to use mysqlslap to perform load test emulation for Azure Database for MySQL – Flexible Server, a fully managed and scalable MySQL service on Azure. I’ll install mysqlslap, configure the connection parameters, run different types of tests, and then analyze the results.

Prerequisites

Before you begin, ensure that the following prerequisites are in place:

An instance of Azure Database for MySQL – Flexible Server. To create one, follow this guidance in this tutorial: Quickstart: Create with Azure portal – Azure Database for MySQL – Flexible Server.
A MySQL client tool that supports mysqlslap – download them from MySQL :: MySQL Community Downloads.
A test database and table on your Azure Database for MySQL Flexible Server instance. Use the following queries to create a test database and table with one million dummy records:

mysql> CREATE DATABASE loadtestdb;

mysql> use loadtestdb;

mysql> CREATE TABLE loadtesttable (
ID INT PRIMARY KEY AUTO_INCREMENT,
Name VARCHAR(255),
Age INT,
Salary DECIMAL(10, 2),
Department VARCHAR(50),
City VARCHAR(100),
Country VARCHAR(100)
);

mysql> INSERT INTO loadtesttable (Name, Age, Salary, Department, City, Country)
SELECT
CONCAT(CHAR(FLOOR(RAND() * 26) + 65), ‘Person’, n),
FLOOR(RAND() * 100) + 18,
ROUND(RAND() * 10000000, 2),
CASE WHEN RAND() < 0.5 THEN ‘IT’ ELSE ‘Sales’ END,
CASE WHEN RAND() < 0.5 THEN ‘New York’ ELSE ‘Los Angeles’ END,
CASE WHEN RAND() < 0.5 THEN ‘USA’ ELSE ‘Canada’ END
FROM (
SELECT
a.N + b.N * 10 + c.N * 100 + d.N * 1000 + e.N * 10000 + f.N * 100000 + 1 AS n
FROM
(SELECT 0 AS N UNION ALL SELECT 1 UNION ALL SELECT 2 UNION ALL SELECT 3 UNION ALL SELECT 4 UNION ALL SELECT 5 UNION ALL SELECT 6 UNION ALL SELECT 7 UNION ALL SELECT 8 UNION ALL SELECT 9) a,
(SELECT 0 AS N UNION ALL SELECT 1 UNION ALL SELECT 2 UNION ALL SELECT 3 UNION ALL SELECT 4 UNION ALL SELECT 5 UNION ALL SELECT 6 UNION ALL SELECT 7 UNION ALL SELECT 8 UNION ALL SELECT 9) b,
(SELECT 0 AS N UNION ALL SELECT 1 UNION ALL SELECT 2 UNION ALL SELECT 3 UNION ALL SELECT 4 UNION ALL SELECT 5 UNION ALL SELECT 6 UNION ALL SELECT 7 UNION ALL SELECT 8 UNION ALL SELECT 9) c,
(SELECT 0 AS N UNION ALL SELECT 1 UNION ALL SELECT 2 UNION ALL SELECT 3 UNION ALL SELECT 4 UNION ALL SELECT 5 UNION ALL SELECT 6 UNION ALL SELECT 7 UNION ALL SELECT 8 UNION ALL SELECT 9) d,
(SELECT 0 AS N UNION ALL SELECT 1 UNION ALL SELECT 2 UNION ALL SELECT 3 UNION ALL SELECT 4 UNION ALL SELECT 5 UNION ALL SELECT 6 UNION ALL SELECT 7 UNION ALL SELECT 8 UNION ALL SELECT 9) e,
(SELECT 0 AS N UNION ALL SELECT 1 UNION ALL SELECT 2 UNION ALL SELECT 3 UNION ALL SELECT 4 UNION ALL SELECT 5 UNION ALL SELECT 6 UNION ALL SELECT 7 UNION ALL SELECT 8 UNION ALL SELECT 9) f
) AS Numbers;

Installing mysqlslap

Instructions for installing mysqlslap on a computer running Windows or Linux appear in the following sections.

Note: If the Azure Cloud Shell is installed, mysqlslap is included automatically.

Windows

To install mysqlslap on a computer running Windows, download the MySQL Installer, run it, and then follow the wizard. Mysqlslap will be installed in the folder C:Program FilesMySQLMySQL Server 8.1bin (assuming the installation is on C:). Alternately, you can download MySQL ZIP Archive and extract mysqlslap from mysql-8.0.36-winx64.zipmysql-8.0.36-winx64bin.

Linux

To install mysqlslap on a computer running Linux, install the MySQL client package (which includes mysqlslap) by running the following command:

sudo apt update
sudo apt install mysql-client
mysqlslap –version

Configuring the connection parameters

To connect to the Azure Database for MySQL – Flexible Server instance, run the following command:

mysqlslap –host=myserver.mysql.database.azure.com –user=myuser –password=mypassword –port=3306 –ssl-mode=REQUIRED

With this command, you can specify the following parameters:

–host: The host name or IP address of your server. You can find it on the Azure portal under the Overview section of your server.
–port: The port number of your server. The default is 3306.
–user: The username to log in to your server. You can use the admin user that you created when you provisioned your server, or any other user that has access to the test database.
–password: The password to log in to your server. You will be prompted to enter it when you run mysqlslap.
–ssl-mode: The SSL mode to use for the connection. You can use REQUIRED, VERIFY_CA, or VERIFY_IDENTITY. The default is REQUIRED. For more information about SSL modes, see MySQL 8.0 : Configuring MySQL to Use Encrypted Connections.

Running different types of tests

The mysqlslap test process includes three stages:

Create schema, table, and optionally any stored programs or data to use for the test. This stage uses a single client connection.
Run the load test. This stage can use many client connections.
Clean up (disconnect, drop table if specified). This stage uses a single client connection.

Use mysqlslap to run different types of tests, such as concurrency tests, stress tests, or benchmark tests. To specify the test parameters, consider the following options:

Option Name

Description

—concurrency

Specifies the number of simultaneous client connections. You can provide a single value or a comma-separated list of values. For example, –concurrency=10 means 10 threads, and –concurrency=10,20,30 means three tests with 10, 20, and 30 threads respectively.

—iterations

Defines the number of times the benchmark test should be repeated. The default is 1.

—number-of-queries

The number of queries to run per thread. The default is 0, which means unlimited.

—query

Specifies the SQL query to be executed during the test. You can provide a single query or multiple queries. For example, –query=”SELECT * FROM testtable” means to run a simple SELECT query.

—create-schema

The name of the database to use for the test. The default is the mysqlslap database.

—create

The statement to create the test table. You can provide a single statement or multiple statements. For example, –create=”CREATE TABLE testtable (id INT)” means to create a simple test table.

—delimiter

Use the –delimiter option to specify a different delimiter, which enables you to specify statements that span multiple lines or place multiple statements on a single line.

—auto-generate-sql

A flag to indicate whether to generate random queries for the test. The default is FALSE. If you set it to TRUE, you can use the following options to control the query generation.

—auto-generate-sql-add-autoincrement

A flag to indicate whether to add an AUTO_INCREMENT column to the test table. The default is FALSE.

—auto-generate-sql-execute-number

The number of queries to generate and execute per thread. The default is 10.

—auto-generate-sql-load-type

The type of queries to generate. You can use MIXED, UPDATE, WRITE, or READ. The default is MIXED.

—auto-generate-sql-unique-query-number

 The number of unique queries to generate. The default is 10.

—auto-generate-sql-unique-write-number

The number of unique queries to generate for write load. The default is 10.

You can also use the –engine option to specify the storage engine to use for the test table. The default is InnoDB. For more information about mysqlslap options, see MySQL 8.0 Reference Manual :: 6.5.8 mysqlslap.

Before running a test with mysqlslap, please ensure to use an empty user database or the default mysqlslap database when using –create or –auto-generate-sql-* option. If the –create or –auto-generate-sql-* option is given, mysqlslap drops the schema at the end of the test run. This means that any existing data in the database will be lost.

Some examples showing how to run different types of tests using mysqlslap follow.

To run a concurrency test with 10, 20, and 30 threads, each executing 100 queries 10 times, run the following command:

mysqlslap –host=server-name.mysql.database.azure.com –port=3306 –user=user-name –password –ssl-mode=REQUIRED –concurrency=10,20,30 –iterations=10 –number-of-queries=100 –query=”SELECT ID, Name, Age, Salary, Department, City, Country FROM loadtesttable WHERE Name like ‘A%’ AND Age BETWEEN 30 AND 40;” –create-schema=loadtestdb –verbose

To run a stress test with 50 threads, each executing the query 20 times, run the following command:

mysqlslap –host=server-name.mysql.database.azure.com –port=3306 –user=user-name –password –ssl-mode=REQUIRED –concurrency=50 –iterations=25 –query=”SELECT ID, Name, Age, Salary, Department, City, Country FROM (SELECT *, ROW_NUMBER() OVER (PARTITION BY Department ORDER BY Salary DESC) AS R FROM loadtesttable) AS ranked WHERE R <= 5;” –create-schema=loadtestdb –verbose

To run a benchmark test with 10 threads, each executing 1000 randomly generated queries with a mixed load type, run the following command:

mysqlslap –host=server-name.mysql.database.azure.com –port=3306 –user=user-name –password –ssl-mode=REQUIRED –concurrency=10 –iterations=1 –number-of-queries=1000 –auto-generate-sql –auto-generate-sql-load-type=MIXED –verbose

Analyzing the results

After running a test, mysqlslap displays the results as output., which includes the:

Average number of seconds to run all queries: The average time it took to run all the queries per thread.
Minimum number of seconds to run all queries: The minimum time it took to run all the queries per thread.
Maximum number of seconds to run all queries: The maximum time it took to run all the queries per thread.
Number of clients running queries: The number of threads that simulated the client load.
Average number of queries per client: The average number of queries that each thread executed.

You can use the –silent option to suppress the verbose output and display only the results. You can also use the –csv option to format the results as comma-separated values, which can easily be imported into a spreadsheet or a database for further analysis.

An example of the results from a concurrency test with 10, 20, and 30 threads, each executing 100 queries 10 times, follows:

Benchmark
Average number of seconds to run all queries: 2.024 seconds
Minimum number of seconds to run all queries: 2.003 seconds
Maximum number of seconds to run all queries: 2.041 seconds
Number of clients running queries: 10
Average number of queries per client: 10

Benchmark
Average number of seconds to run all queries: 2.070 seconds
Minimum number of seconds to run all queries: 2.022 seconds
Maximum number of seconds to run all queries: 2.228 seconds
Number of clients running queries: 20
Average number of queries per client: 5

Benchmark
Average number of seconds to run all queries: 1.885 seconds
Minimum number of seconds to run all queries: 1.849 seconds
Maximum number of seconds to run all queries: 2.021 seconds
Number of clients running queries: 30
Average number of queries per client: 3

You can use the results to compare the performance of an Azure Database for MySQL – Flexible Server instance under different load scenarios, and to identify any potential bottlenecks or issues. You can also use the results to tune the server configuration, such as the number of connections, the buffer pool size, the query cache size, or the index statistics.

Best practices

When you’re using mysqlslap to perform load test emulation for an Azure Database for MySQL – Flexible Server instance, consider the following best practices.

Before using the mysqlslap utility on your production environment, test the mysqlslap utility thoroughly in your lowest environment against a test database.
Define benchmarking scenarios that closely resemble your production environment.
Use realistic datasets and queries representative of your actual workload.
Adjust benchmarking parameters such as concurrency, iterations, and query complexity to match your workload characteristics.
Test different combinations of parameters to understand their impact on performance.
Analyze results carefully and consider multiple metrics for performance evaluation.
Monitor system resources (CPU, memory, disk I/O) during benchmark tests to identify any resource bottlenecks.
Repeat benchmark tests multiple times to validate results and ensure consistency.

Conclusion

In this post, I’ve described how to use mysqlslap to perform load test emulation for an Azure Database for MySQL – Flexible Server instance. I’ve described how to install mysqlslap, configure the connection parameters, run different types of tests, and analyze the results. Be sure to use mysqlslap to simulate client load and measure the performance of your MySQL flexible server, as well as to optimize your server configuration and query performance.

If you have any questions about the detail provided above, please leave a comment below or email us at AskAzureDBforMySQL@service.microsoft.com. Thank you!

References

For more information about using mysqlslap, in the MySQL documentation, see https://dev.mysql.com/doc/refman/8.0/en/mysqlslap.html

​Microsoft Tech Community – Latest Blogs –Read More 

Note: this is the second of a four-part blog series that explores the complexities of securing multiple clouds and the limitations of traditional Security Information and Event Management (SIEM) tools.

With the first post, we discussed the importance of adopting a multi-cloud approach to Observability, centralizing in a single SIEM all the events generated by your infrastructure to enable a more comprehensive analysis of potential security incidents by correlating events independently from their origin. We also hinted to the complexity of such endeavor.

You can read the first post here: Securing the Clouds: Navigating Multi-Cloud Security with Advanced SIEM Strategies – Microsoft Community Hub

With this new post, we focus on a different topic: the importance of adopting a threat-based approach. In the process, we discuss how this can be achieved and provide you with a few practical ideas you can apply to your scenarios.

The Threat-Based Approach

The threat-based approach for creating use cases consists in the identification of potential attacks to the system, considering each Cloud environment, on-prem environment, and then how they interrelate and interact. You then derive attack uses cases which drive the definition of the logic to identify those attacks and then trigger remediation activities. Those potential attacks are also known as threats or, more precisely, as threat events.

The threat-based approach is not the only possibility. Actually, the most common approaches are vulnerability-based. With them, the focus is on the identification of vulnerabilities like the infamous Log4Shell, and consequently on indicators which may identify attacks in progress.

Vulnerability-Based Approaches and How They Compare

Vulnerability-driven approaches have various shortcomings. For instance, they tend to grow the detection capabilities to respond to known vulnerabilities, and as a reaction to successful attacks. Therefore, they are designed to prevent such attacks from occurring. Conversely, threat driven approaches are based on the understanding of how attacks may happen, and are designed to detect them independently from the presence of specific vulnerabilities. We have chosen to adopt a threat driven approach to improve the detection and response capabilities and to reduce the impact of security incidents and response.

Another advantage of the threat-driven approach is that it is often independent from the existence of specific vulnerabilities. For example, you can analyze the possibility for an attacker to inject code in its requests leveraging some vulnerability to execute such code, without referring to a specific vulnerability. This allows you to design detection mechanisms that are vulnerability independent. Therefore, they would apply to many similar vulnerabilities, including yet undisclosed zero-day vulnerabilities.

A threat-driven approach is a proactive and strategic way of analyzing a complex infrastructure for identifying This approach is much more effective of the corresponding ones based on vulnerability analysis because it takes in account the exploitability of the vulnerabilities. In other words, you determine what a malicious actor can do to compromise your system. This is much more effective than adopting a vulnerability-driven approach because it allows you to discard vulnerabilities that do not matter because they are hardly exploitable.

How to Apply the Threat-Based Approach

By adopting a threat-driven approach we started with the following steps:

Analyze the organization threat landscape focusing on factors like geographical location, industry, externally exposed services, and potentially much more.
Leverage the organization’s threat intelligence to gain broader understanding of the threat landscape, and the specific threats and attacks that target them.
Identify and prioritize the most impactful threats and attack vectors that target the organization’s assets, operations, and objectives, using your telemetry.
Assess and understand the capabilities, tactics, techniques, and procedures of the threat actors and their motivations and goals.
Monitor and evaluate the effectiveness and performance of the security controls and countermeasures and adjust them over time as the threat landscape evolves.

To apply a threat-driven approach, organizations need to incorporate threat analysis and threat intelligence across their systems development and operational processes.

A Practical Example of Threat-Based Approach

Now that we have understood what the threat-driven approach is, it is time to get a few ideas about how you can implement it in your organization.

We can consider as an example a Multi-Cloud organization. It is not uncommon for those organizations to adopt identity and security solutions from various vendors. This complicates integration and soars the risk of supply chain attacks due to the lack of comprehensive visibility and increased complexity. To address this situation, it is best to adopt a structured approach based on the following steps:

Get a good understanding of the various environments, focusing your attention on their components and on how they interact with each other.
Perform a risk analysis on the said environments to identify threats, monitoring capabilities included with each service, and identifying eventual gaps to be covered with additional events. This could be done by applying some lightweight threat modeling.
Evaluate the typical attack scenarios seen by the organization towards the systems in scope and ensure that they are represented within the threat analysis. This may include an analysis of the specific threat landscape for the organization.

The “threat modeling” specified in the second point is a security process to understand security threats to a system, determine risks from those threats, and establish appropriate mitigations. There are various ways to perform threat modeling, and all of them are well represented by the Threat Modeling Manifesto. Microsoft has developed one of the first threat modeling processes, called Microsoft STRIDE Threat Modeling. This approach has evolved over the years and is continuously evolving. If you want to learn more, please go to Microsoft Security Development Lifecycle Threat Modelling.

Creating the Use Cases

The three steps defined above define the threats for the system. This represents the first phase of our journey. The next one consists of the definition of Use Cases. You will often have a Use Case for each Threat, but this is not an absolute rule. In various situations you will want to cover more threats with a single Use Case. Nevertheless, each Use Cases describes the associated threats as a single story and identifies the events necessary to detect such attacks and where they can be found. The Use Case typically also contains the definition of the actions that can be performed for controlling the risk, and the conditions under which they are triggered. Those actions will be both manual and automated.

Responding to Attacks

It is very important to define automated activities that are executed when a potential attack is detected. This allows you to respond to attacks faster than you could if you rely only on manual response.

Time is a critical factor in the realm of cybersecurity. Let’s delve into why swift responses matter:

When a cyberattack occurs, a rapid response allows organizations to identify and contain the breach promptly. By doing so, they prevent the attack from spreading or escalating into a larger incident. This quick action minimizes the damage inflicted on systems, data, and operations.
Fast response enables organizations to restore normal operations swiftly. By minimizing downtime, they mitigate financial losses and maintain business continuity.
The cybersecurity landscape involves a constant race between defenders and attackers. Cybercriminals leverage evolving tools, tactics, and procedures, including zero-day exploits. While it takes an independent cybercriminal around 9.5 hours to gain illicit access to a target, defenders must act even faster to thwart such attempts. See The Importance Of Time And Speed In Cybersecurity (forbes.com).

While automation plays a major role in reducing the time required to respond to attacks, manual remediation activities are still essential. The problem is that automated actions cannot be too drastic: it’s fine to disable an IP address or an account that seems to attack the organization, but you will not want to have some automated procedure to take down the whole infrastructure automatically, even if you detect a possible data exfiltration.

Ultimately, the production of the Use Cases is instrumental in the creation of the rules in your SIEM system to detect attacks, and to configure your SOAR system for automatically respond to them.

Conclusions

Complex Multi-Cloud environments represent a significant challenge when you must create a monitoring infrastructure. Yet, achieving a comprehensive view of what happens in your organization is more important than ever. A structured approach like the Threat-Based Approach described in this post may help you to conquer complexity and get the results your organization needs.

Nevertheless, implementing a Threat-Based Approach is not the end. Organizations face new attacks every day. New software is acquired, extending the attack surface for the Organization. And new vulnerabilities are regularly found. For these reasons, it is essential to adopt a continuous improvement approach. The threat assessment must be regularly reiterated, and new Use Cases must be created. This leads to updating the existing rules in the SIEM or SOAR. If you do so, your Organization will gradually but surely improve its security posture, and will eventually become one of the main tools to guarantee your Business’ security.

Future posts in this series will cover the following topics:  

How Microsoft has implemented its security solutions across Azure, Oracle, AWS, and on-premises environments, thus enabling a unified and comprehensive defense against threats, for any enterprise
Key benefits and outcome examples for some of our multi-cloud security projects, including improved detection capabilities, enhanced visibility across enterprise, efficiency, and cost savings. 

​Microsoft Tech Community – Latest Blogs –Read More 

This blog focuses on good practices for monitoring Azure Open AI limits and quotas. With the growing interest and application of Generative AI, Open AI models have emerged as pioneers in this transformative era. To maintain consistent and predictable performance for all users, these models impose certain limits and quotas. For Independent Software Vendors (ISVs) and Digital Natives utilizing these models, understanding these limits and establishing efficient monitoring strategies is paramount to ensures a good customer experience to the end-users of their products and services. This blog seeks to provide a comprehensive understanding of these monitoring strategies, thereby enabling ISVs and Digital Natives to optimally leverage AI technologies for their respective customer bases.

Understanding Limits and Quotas

Azure OpenAI’s quota feature enables assignment of rate limits to your deployments, up-to a global limit called your “quota”. Quota is assigned to your subscription on a per-region, per-model basis in units of **Tokens-per-Minute** (TPM). Your subscription is onboarded with a default quota for most models.

Refer to this document for default TPM values. You can allocate TPM among deployments until reaching quota. If you exceed a model’s TPM limit in a region, you can reassign quota among deployments or request a quota increase. Alternatively, if viable, consider creating a deployment in a new Azure region in the same geography as the existing one.

For example, with a 240,000 TPM quota for GPT-35-Turbo in East US, you could create one deployment of 240K TPM, two of 120K TPM each, or multiple deployments adding up to less than 240K TPM in that region.

TPM rate limits are based on the maximum tokens **estimated** to be processed when the request is received. It is different than the token count used for billing, which is computed after all processing is completed. Azure OpenAI calculates a max processed-token count per request using:

– Prompt text and count
– The max_tokens setting
– The best_of setting

This estimated count is added to a running token count of all requests, which resets every minute. A 429 response code is returned once the TPM rate limit is reached within the minute.

A **Requests-Per-Minute** (RPM) rate limit is also enforced. It is set proportionally to the TPM assignment at a ratio of 6 RPM per 1000 TPM. If requests aren’t evenly distributed over a minute, a 429 response may be received. Azure OpenAI Service evaluates incoming requests’ rate over a short period, typically 1 or 10 seconds, and issues a 429 response if requests surpass the RPM limit. For example, if the service monitors with a 1-second interval, a 600-RPM deployment would be throttled if more than 10 requests are received per second.

In addition to the standard quota, there is also a provisioned throughput capability, or PTU. It is useful to think of the standard quota as a serverless mode, where your requests are served from a pool of resources and no capacity is reserved for you, hence the overall latency could vary. In contrast, with a provisioned throughput capability, you specify the amount of throughput you require for your application. The service then provisions the necessary compute and ensures it is ready for you. This gives you a more predictable performance and stable max latency. For high throughput workloads this may provide cost savings versus the token-based consumption. At the time of the writing, provisioned throughput units are not available by default. For more details about it, contact your Microsoft Account team.

There is also a limit of 30 Azure OpenAI resource instances per region. For an exhaustive and up-to-date list of quotas and limits please check this document. It is important to plan ahead on how you will manage and segregate tenant data and traffic in order to ensure reliable performance and optimal costs. Please check the Azure Open AI service specific guidance for considerations and strategies pertinent to multitenant solutions.

Choosing between tokens-per-minute and provisioned throughput models

To choose effectively between TPM and PTU you need to understand that there are minimum PTUs per deployment required. If your current usage is above the requirement and expected to grow, it might be more economically feasible to purchase provisioned capacity. In high token usage scenarios, this provides a lower per token price and stable max latency. It is important to understand that with PTUs, you are isolated and protected from the noisy neighbor problem of a SaaS application with shared resources. However, you can still experience higher than average latency caused by other factors, such as the total load you send to the service, length of the prompt and response, etc.

The table below shows the minimum TPMs per model type and an approximate relation to TPMs:


Source: https://github.com/Azure/aoai-apim

Effective Monitoring Techniques

Now that we understand better the limits and quotas of the service, let’s discuss how to effectively monitor usage and set up alerts to be notified and take action when you reach the limits and quotas assigned.

Azure OpenAI service has metrics and logs available as part of the Azure Monitor capabilities. Metrics are available out of the box, at no additional cost. By default, a history of 30 days is kept. If you need to keep these metrics for longer, or route to a different destination, you can do so by enabling it in the Diagnostic settings.

Metrics are grouped into four categories:

– HTTP Requests dimensions: Model Name, Model Version, Deployment, Status Code, Stream Type, and Operation.
– Tokens-Based Usage: Active tokens, Generated Completions Tokens, Processed Inference and Prompt Tokens.
– PTU Utilization dimensions: Model Name, Model Version, Deployment, and Stream Type.
– Fine-tuning: Training Hours by Deployment and Training Hours by Model Name.

Additionally, each API response header contains the RateLimit-Global-Remaining and RateLimit-Global-Reset. And the response body contains a usage section with the prompt tokens, completion tokens, and total tokens values that shows the billing tokens per request.

The available logs in Azure OpenAI are Audit logs, Request and Response logs, and Trace Logs. Once you enable these through the Diagnostic settings, you can send these to a Log Analytics workspace, Storage account, Event Hub, or a partner solution. Keep in mind that using diagnostic settings and sending data to Azure Monitor Logs has other costs associated with it. For more information, see Azure Monitor Logs cost calculations and options.

My colleagues created an Azure Monitor Workbook that serves as a great baseline to start monitoring your Azure Open AI service logs and metrics.

Optimization Recommendations

Use LLMs for what they are best at – natural language understanding and fluent language generation. This means understanding that LLMs are boxes trying to predict the most likely next token and just because you could use an LLM for a task, it doesn’t necessarily make it the most optimal tool for it.

1. Always start with can this be done in code? Are there existing libraries, tools, patterns that can perform the task? If yes, use those. These will probably be more performant and cost less.
Examples: use Azure AI Language service for key phrase extraction instead of the LLM; use standard libraries to do math operations, data aggregation, etc.

2. Control the size of the input prompt (e.g. set a limit on the user input field; in RAG, depending on scenario, restrict the number of relevant chunks sent to the LLM) and completion (with max_tokens and best_of).

3. Call the GPT models as few times as possible. Ensure you gather all the data you need to generate an optimal response, and only then call the model.

4. Use the cheapest model that gets the task done. This could mean using GPT 3.5 instead of GPT 4 for tasks where the cheapest model performs at an acceptable level.

Prevention and Response Strategies for Limit Exceeding

Here are some best practices and strategies to avoid rate limiting errors in a tokens per minute, i.e. Pay-As-You-Go model:

– Use minimum feasible values for max_tokens and best_of in your scenario. For instance, don’t set a high max-tokens if expecting small responses.
– Manage your quota to allocate more TPM to high-traffic deployments and less to those with limited needs.
– Avoid sharp changes in the workload. Increase the workload gradually.
– Test different load increase patterns.
– Check the size of prompts against the model limits before sending the request to the Azure Open AI service. For example, for GPT-4 (8k), a max request token limit of 8,192 is supported. If your prompt is 10K in size, then this will fail, and also any subsequent retries would fail as well, consuming your quota.
– retrying with exponential backoff – in practice it means performing a short sleep when a rate limit error is hit, then retrying the unsuccessful request. If the request is still unsuccessful, the sleep length is increased and the process is repeated. Note that unsuccessful requests contribute to your per-minute limit, so continuously resending a request won’t work. This strategy is useful for real-time requests from users.
– batching requests – If you’re hitting the limit on requests per minute, but have headroom on tokens per minute, you can increase your throughput by batching multiple tasks into each request. This will allow you to process more tokens per minute, especially with the smaller models.
– when handling batch processing, maximizing throughput matters more than latency. Proactively adding delay between batch requests can help. E.g., if your rate limit 20 requests per minute, add a delay of 3–6 seconds to each request. This can help you operate near the rate limit ceiling without hitting it and incurring wasted requests.

For more details on these strategies and an example of a parallel processing script, please see this notebook and documentation from Azure Open AI.

If your workload is particularly sensitive to latency and cannot tolerate latency spikes, you can consider implementing a mechanism that checks the latency of Azure Open AI in different Azure regions and send requests to the region with the smallest latency. You can group regions into geographies, like Americas, EMEA and Asia, and perform these checks on a per geography basis. This should also account for any compliance regulation and data residency requirements. For a more detailed walkthrough of this strategy, please check this blog.

In Azure, API Management (APIM) service can help you implement some of these best practices and strategies. APIM supports queueing, rate throttling, error handling, managing user quotas, as well as distributing requests to different Azure Open AI instances, potentially located in different regions to implement the pattern described above.

Conclusion

In conclusion, understanding the limits, quotas, and optimization techniques for Azure Open AI is crucial for effectively utilizing the service and achieving optimal performance and cost efficiency. By carefully monitoring usage, setting up alerts, and implementing prevention and response strategies for limit exceeding, you can ensure reliable performance and avoid unnecessary disruptions.

The insights and recommendations provided in this document serve as a valuable guide to help you make informed decisions and optimize your Azure Open AI use-cases. By following these best practices, such as leveraging existing libraries and tools, controlling input prompt size, minimizing API calls, and using the most cost-effective models, you can maximize the value and efficiency of your AI applications.

Remember to plan ahead, allocate resources wisely, and continuously monitor and adjust your usage based on the metrics and logs available through Azure Monitor. By doing so, you can proactively address any potential issues, avoid rate limiting errors, and deliver a seamless and responsive experience to your users.

​Microsoft Tech Community – Latest Blogs –Read More 

Learn about SQL Server and Azure SQL Database security fundamentals you won’t want to miss.

Resources:

Azure Security Podcast

​Microsoft Tech Community – Latest Blogs –Read More 

The Microsoft 365 Copilot Adoption Accelerator engagement is crafted to ensure the seamless adoption of Copilot for Microsoft 365.

This engagement comprises three key phases: Readiness, Build the Plan, and Drive Adoption. It is recommended to undertake the Adoption Accelerator after completing the Copilot for Microsoft 365 engagement, wherein high-value scenarios and the technical and organizational baseline are identified. The Adoption Accelerator Engagement will specifically target these high-value scenarios. 

The adoption process should involve key stakeholders such as Adoption Managers, Business Decision Makers, End-User Support, and Champions. Responsibility for sustained success should be effectively transitioned during the adoption process. 

Click here for more information 

​Microsoft Tech Community – Latest Blogs –Read More 

Trevor Noah, Microsoft Chief Questions Officer and renowned comedian, author, and former host The Daily Show, joined the keynote stage at the Global Nonprofit Leaders Summit with Kate Behncken, Global Head of Microsoft Philanthropies, for a conversation about social impact.

From his first childhood encounter with a PC (it was a Pentium 386!) to working with Microsoft AI for Good, Trevor meets the opportunity of technology with natural curiosity and optimism that inspires everyone to find ways to use AI to build equity, fairness, and security for people around the world.

He shares examples from the AI for Good projects he’s featured on his series “The Prompt” and talks in depth about how AI is creating a critical moment for expanding education and opportunities in developing countries. Then with his inimitable humor, he somehow manages to include buffalo wings as an example of how we should always ask ourselves, “What if I’m wrong?” 

What did you learn from Trevor Noah’s insights? What are some examples of when you’ve asked yourself, “What if I’m wrong?”

​Microsoft Tech Community – Latest Blogs –Read More 

Today, we are excited to announce the expansion of Copilot for Microsoft 365 with the general availability of new offerings for sales, services, and education—providing new ways for customers and partners to embrace AI to achieve their goals and significantly enhance how they work.  

Copilot for Sales is an AI assistant for sales professionals that brings together the power of generative AI through Copilot for Microsoft 365, Copilot Studio, and data from any CRM system to accelerate productivity, keep data fresh, unlock seller-specific insights, and help sellers personalize customer interactions – all leading to closure of more deals.  

Copilot for Service modernizes existing service solutions with generative AI to enhance customer experiences and boost agent productivity. It infuses AI into the contact center to accelerate time to production with point-and-click setup, direct access within major service vendors (including Salesforce, ServiceNow, and Zendesk), and connection to public websites, SharePoint, knowledgebase articles, and offline files. Agents can ask questions in natural language and answers are delivered in the tools they use every day—Outlook, Teams, Word, and others.  

Learn more about Copilot for Sales and Copilot for Service announcements. 

In addition, Copilot for Microsoft 365 is now also generally available for education customers to purchase for their faculty users through the Cloud Solution Provider (CSP) program.   

Read the announcement  
Plan to join us on Wednesday, March 6, 2024, for the Reimagine education event: the future of AI in Education for an opportunity to deep dive on Copilot for Education.  
Read more in our blog about specific opportunities for education partners with Copilot for Microsoft 365. 

Additional opportunities for all partners  

Register for the Copilot partner incentives overview webinar – February 28/29 to learn about the latest priorities, strategy, and earning opportunities for Copilot & AI.  
Sign up for a Copilot for Microsoft 365 pre-sales and technical bootcamp  
Join the conversation on our Copilot for Microsoft 365 community

​Microsoft Tech Community – Latest Blogs –Read More