benchmarking Archives

AppWizard

April 11, 2026

Cities Skylines 2’s performance is still struggling, and the devs need your help

Cities Skylines 2 has faced performance issues and limited content since its launch. Iceflake Studios is working on improvements, collaborating with Unity to address performance challenges. The game is CPU-intensive, with unique challenges due to its dynamic nature. Key areas for enhancement include rendering performance, simulation performance, and pathfinding. Recent changes have simplified citizen modeling and improved pathing, leading to a better gameplay experience. The game's Steam rating has shifted from 'Mixed' to 'Mostly Positive.' Iceflake has developed a benchmarking tool to collect performance data, which will be integrated into an upcoming patch.

AppWizard

April 9, 2026

Google updates best AI models for coding Android apps, Gemini & GPT 5.4 at the top

The "Android Bench," Google's benchmark for evaluating AI models in Android app development, has been updated, with OpenAI's GPT 5.4 and GPT 5.3 Codex now sharing the top ranking with Gemini. The benchmark evaluates models based on criteria such as compatibility with Jetpack Compose, use of Coroutines and Flows, and integration with Room and Hilt. The latest rankings are as follows: 1. GPT 5.4: 72.4% 2. Gemini 3.1 Pro Preview: 72.4% 3. GPT 5.3-Codex: 67.7% 4. Claude Opus 4.6: 66.6% 5. GPT-5.2 Codex: 62.5% 6. Claude Opus 4.5: 61.9% 7. Gemini 3 Pro Preview: 60.4% 8. Claude Sonnet 4.6: 58.4% 9. Claude Sonnet 4.5: 54.2% 10. Gemini 3 Flash Preview: 42% 11. Gemini 2.5 Flash: 16.1% The rankings have not changed since the initial assessment in late February, and the latest models were evaluated in mid-March. The findings should be interpreted cautiously, as real-world performance may vary based on specific workflows and project requirements.

Tech Optimizer

April 5, 2026

Linux 7.0 Cuts PostgreSQL Performance in Half

An AWS engineer reported a significant drop in PostgreSQL throughput on Linux 7.0, with performance reduced to approximately half of its previous capability. Benchmark tests showed that the removal of the PREEMPT_NONE scheduling option was the main cause of this regression. On a 96-vCPU Graviton4 instance, throughput measured at just 0.51x compared to earlier kernel versions. Salvatore Dipietro from Amazon/AWS conducted benchmarking analysis of PostgreSQL 17, revealing that Linux 7.0 delivered only 0.51x the throughput of its predecessors. The root cause was traced to kernel commit 7dadeaa6e851, which eliminated PREEMPT_NONE as the default option, leading to increased contention due to the new PREEMPT_LAZY model. Profiling data indicated that 55% of CPU time is consumed by spinning in PostgreSQL’s spinlock, causing significant performance degradation. When a revert patch was applied, throughput rebounded to 1.94x the baseline. The decision to restrict preemption modes in Linux 7.0 aimed to address issues within the kernel's scheduling model. Dipietro proposed a patch to restore PREEMPT_NONE, but kernel developers suggested PostgreSQL adopt the rseq time slice extension instead. Database operators running PostgreSQL on Linux face potential performance reductions with the upgrade to Linux 7.0.

Tech Optimizer

April 1, 2026

EDB Postgres® AI Delivers Superior Predictability vs. Cloud Data Warehouses in High-Concurrency Benchmark, Unveils Q1 Platform Updates to Power the Agentic AI Era

Independent benchmarking by McKnight Consulting Group shows that EDB Postgres AI for WarehousePG provides significant cost efficiency and performance consistency, with organizations potentially saving up to 58% in total cost of ownership compared to leading cloud data warehouse solutions. The evaluation compared EDB PG AI against competitors like Snowflake, Databricks, Amazon Redshift, and Hive on Apache Iceberg using a 10TB extended TPC-DS dataset, focusing on high-concurrency mixed workloads. Key findings include: - EDB PG AI demonstrated unmatched cost efficiency, with an annual cost of ,886 compared to Snowflake’s ,953 for a multi-cluster setup. - It exhibited superior concurrency handling, with the lowest performance slowdown of 2.7x when scaling from one to five concurrent users, outperforming Snowflake (3.9x), Redshift (4.0x), and Databricks (4.1x). - EDB PG AI's core-based, capacity-pricing model eliminates unpredictable pricing fluctuations associated with consumption-based models. EDB announced Q1 2026 platform updates, including: - GPU-Accelerated Analytics for 50–100x faster analytics on large datasets. - Enhanced Agent Studio for quicker AI agent development and deployment. - Upgraded Vector Engine for improved indexing speed and efficiency. - WarehousePG Enterprise Manager for simplified management of MPP workloads. - Agentic Database Management with a native chatbot for natural language database management. - Certification as a mission-critical data layer for the Red Hat Ansible Automation Platform.

Tech Optimizer

March 29, 2026

PostgreSQL cost model: estimating the cost of flushing temp buffers to disk

The project focuses on enabling temporary tables in PostgreSQL's parallel query execution by benchmarking sequential writes and reads of temporary buffers. Functions were introduced to measure buffer flush operations, and tests were conducted to estimate the cost of flushing temporary table buffer pages to disk. Findings indicate that sequential writes are about 30% slower than reads, leading to a proposed cost model. Temporary tables have limited parallel usage due to their function as relational variables and the complexity of parallel workers accessing the leader process's local state. Two enterprise PostgreSQL distributions already support parallel operations on temporary tables, and discussions on logical replication of DDL highlight the need for improved tooling for temporary tables. A significant challenge is that temporary buffer pages are local to the leader process, and if not flushed to disk, parallel workers cannot access them. A comment from 2015 suggests flushing the leader's temporary buffers to disk before parallel operations to allow workers to scan the table in parallel. However, this raises concerns about the cost of writing buffers versus the benefits of parallelism. The benchmarking toolkit includes modifications to PostgreSQL to facilitate measurement of local buffers. Two key commits added statistics infrastructure to track allocated and dirty local buffers and UI functions to manage buffer operations. The methodology involves creating a temporary table, flushing buffers, measuring I/O, and evaluating scan overhead. The tests reveal that the cost of writing a temporary table approximates the cost of a sequential page, leading to a proposed formula for flushing costs. The results show that larger buffer sizes lead to higher write overhead and increased variability. A comprehensive formula for pre-flushing temp buffers is suggested, factoring in the number of dirtied and allocated local buffers. Next steps include adding planner flags for temporary objects, implementing executor support for worker access to temporary table storage, introducing a temp-buffer flush operation, and integrating the cost model into the planner for parallel scans of temporary tables.

Winsage

March 19, 2026

Despite hardware limits, Parallels supports running Windows on MacBook Neo

Apple's MacBook Neo is competitively priced at [openai_gpt model="gpt-4o-mini" prompt="Summarize the content and extract only the fact described in the text bellow. The summary shall NOT include a title, introduction and conclusion. Text: Performance Insights of Apple’s MacBook Neo Apple’s latest offering, the MacBook Neo, has garnered attention for its competitive pricing at 0. While it impresses with its affordability, the device does come with certain trade-offs, particularly concerning its A18 Pro processor. In our assessment, the Neo excels in handling basic computing tasks, yet it falls short when faced with demanding workloads that require enhanced CPU and GPU capabilities, as well as additional RAM. For users whose needs extend beyond the basics, the MacBook Air remains the superior option. Despite its limitations, the MacBook Neo proves to be a capable machine for running Windows through Parallels Desktop virtualization software. Parallels has conducted thorough testing and benchmarking, concluding that the Neo is well-suited for “lightweight computing and everyday productivity.” Users can comfortably engage in document editing and utilize web-based applications while running Windows 11. According to Parallels, the MacBook Neo’s commendable single-core CPU performance contributes to a user experience that feels “quick and responsive.” This responsiveness is particularly evident when operating multiple Windows-only software applications. Notable programs such as QuickBooks Desktop, Microsoft Office, and various engineering and data tools—including AutoCAD LT and MATLAB—run smoothly on the Neo. Additionally, it supports specialized educational software that lacks a Mac equivalent. In comparative testing, the Neo demonstrated a single-core CPU performance that was approximately 20 percent faster than that of a Core Ultra 5 235U chip found in the Dell Pro 14 laptop. This performance metric underscores the Neo’s potential for users who prioritize efficiency in their everyday computing tasks." max_tokens="3500" temperature="0.3" top_p="1.0" best_of="1" presence_penalty="0.1" frequency_penalty="frequency_penalty"] and excels in basic computing tasks but struggles with demanding workloads due to limitations in its A18 Pro processor. It is suitable for running Windows through Parallels Desktop, performing well in lightweight computing and everyday productivity tasks like document editing and web applications. The Neo has commendable single-core CPU performance, which is about 20 percent faster than the Core Ultra 5 235U chip in the Dell Pro 14 laptop, making it efficient for users focused on everyday computing. It can run various programs, including QuickBooks Desktop, Microsoft Office, AutoCAD LT, and MATLAB, as well as specialized educational software without a Mac equivalent.

AppWizard

March 6, 2026

Google will now show which AI models are best at building Android apps

Google has introduced Android Bench, a tool for assessing AI model performance in Android app development. The top performer is Gemini 3.1 Pro, scoring 72.2%, followed by Claude Opus 4.6 at 66.6% and GPT 5.2 Codex at 62.5%. The benchmark evaluates models through real-world Android coding challenges with task completion rates between 16% and 72%. Google aims to facilitate the creation of Android applications from user prompts and has made the benchmark's methodology and tools available on GitHub.

Tech Optimizer

February 20, 2026

Linux 7.0 Shows Significant PostgreSQL Performance Gains On AMD EPYC Review

Initial benchmarking of the Linux 7.0 kernel on the Core Ultra X7 "Panther Lake" platform revealed performance regressions. In contrast, testing on an AMD EPYC Turin server showed no regressions and highlighted significant performance enhancements for PostgreSQL database operations. The benchmarks compared Linux 6.19 and Linux 7.0 Git, using an AMD EPYC 9755 single-socket setup on a Gigabyte MZ33-AR1 server. The upgrade to Linux 7.0 resulted in modest improvements for CockroachDB and notable enhancements in PostgreSQL 18.1 for read and write operations. Performance for in-memory databases like Memcached and Pogocache remained unchanged, while slight improvements were observed for the Nginx HTTPS web server and the Open Image Denoise library. The Panther Lake tests had shown increased context switching times, which were not replicated in the AMD EPYC Turin tests. Both platforms indicated enhancements in kernel message passing performance and improvements in socket activity and pthread performance. Ongoing benchmarking will continue as the Linux 7.0 merge window approaches its conclusion.

Tech Optimizer

January 8, 2026

Faster Is Not Always Better: Choosing the Right PostgreSQL Insert Strategy in Python (+Benchmarks)

Inserting 2 million records per second into Postgres is achievable. The analysis explores five methods for inserting data into Postgres using Python, focusing on trade-offs in abstraction, safety, convenience, and performance rather than just speed. High-volume insert workloads are common in scenarios like loading records, syncing data, backfilling analytics tables, and ingesting events. Minor inefficiencies can lead to significant performance impacts. To interact with Postgres, the psycopg3 driver is used alongside SQLAlchemy, which provides two layers: Core and ORM. Psycopg3 is a low-level driver requiring manual SQL management, while SQLAlchemy Core offers a SQL abstraction, and the ORM maps Python classes to database tables, enhancing productivity but introducing overhead. Benchmarking involves measuring only the time spent transferring data from Python to Postgres, ensuring a fair comparison among methods. The fastest method may not always be the best due to maintenance costs, correctness guarantees, and cognitive load. The right insertion strategy depends on the existing data structure rather than just row count. The ORM is suited for CRUD-heavy applications, Core for data ingestion and analytics, and the Driver for maximum throughput in extensive writes. Performance issues can arise from mismatching abstractions, and reverting to a lower level may enhance performance. A guideline for choosing methods is: - Use ORM for applications prioritizing correctness and productivity. - Use Core for data movement or transformation balancing safety and speed. - Use Driver for pushing performance limits with raw power and full responsibility.

Tech Optimizer

January 6, 2026

How to speed up mass data inserts in PostgreSQL when using Spring

A common task in enterprise systems involves loading large volumes of data into PostgreSQL, which can be slow if using a loop in Java that calls the save() method for each record. Initial tests show that a basic insert operation takes over 10 minutes. Hibernate retains every object in the session-level cache until the transaction concludes, leading to high memory consumption. Enabling batching with the jdbc.batch_size setting improves performance by about 23%, but high memory usage persists. Application-side caching of IDs is faster than database-side caching, achieving a 30% performance gain. Using saveAll() results in minimal improvement as it still calls save() in a loop. Enabling reWriteBatchedInserts in the JDBC driver allows for multi-row inserts, reducing operation time to 4 minutes and 37 seconds. Clearing the session-level cache with EntityManager further optimizes memory usage, resulting in an 18-second time gain. Setting order_inserts=true enhances batching efficiency. With all optimizations, insertion time is reduced to 4 minutes and 19 seconds. A custom data insertion layer can be created to generate INSERT statements, with reflection being a straightforward but not optimal method. Prepared statements improve performance and security, with an efficient batch size of 5,000 records. PostgreSQL's COPY command facilitates faster bulk data transfers than multi-row inserts. Integrating COPY with Spring transactions enhances performance, and parallelizing the insertion process with a ThreadPoolExecutor significantly improves speed and memory usage.