Unlocking Peak Performance: Optimizing Dual Xeon Servers on Ubuntu 24.04

Achieving maximum throughput on high-end bare metal infrastructure requires precise OS-level tuning. System administrators must move beyond default configurations to fully leverage modern hardware capabilities.

This guide focuses on a specific enterprise-grade setup featuring Intel Xeon Gold 6148 processors. The goal is to extract every ounce of computational power from this robust architecture running Ubuntu 24.04.

Key Hardware Specifications and Baseline

Understanding the underlying hardware is the first step toward effective optimization. The current configuration represents a powerful mid-to-high-tier enterprise setup.

• CPU: 2x Intel(R) Xeon(R) Gold 6148 @ 2.40GHz (Skylake-SP architecture)
• Memory: 256GB DDR4 ECC RAM
• Storage: 2TB NVMe M.2 SSDs for ultra-low latency I/O
• Network: 1Gbps dedicated bandwidth with 300TB monthly data allowance
• OS: Ubuntu 24.04 LTS (Noble Numbat)
• Deployment: Two identical bare metal servers

These specifications provide a solid foundation for heavy workloads, including database clusters, container orchestration, or AI inference tasks. However, raw power alone does not guarantee efficiency.

Kernel Tuning and CPU Frequency Scaling

The Linux kernel manages hardware resources, making it the primary target for optimization. Default settings often prioritize energy savings over raw performance.

Disabling Power Saving Modes

Modern CPUs like the Xeon Gold series include aggressive power management features. While useful in cloud environments, these features introduce latency spikes in bare metal deployments.

Administrators should disable CPU frequency scaling. Set the governor to 'performance' mode to ensure the CPU runs at its maximum turbo frequency consistently. This prevents the processor from downclocking during idle periods.

Additionally, disable Intel SpeedStep and C-states in the BIOS/UEFI if possible. These features reduce power consumption but add microsecond-level delays when waking cores. For high-throughput applications, consistent clock speeds are superior to variable frequencies.

NUMA Awareness Configuration

The Xeon Gold 6148 utilizes Non-Uniform Memory Access (NUMA) architecture. Each CPU socket has local memory attached directly to it. Accessing remote memory across the QPI link incurs significant latency penalties.

Applications must be pinned to specific NUMA nodes. Use tools like numactl to bind processes to the correct CPU cores and memory regions. This ensures that data stays close to the processing unit, reducing bus contention.

Proper NUMA binding can improve memory-intensive application performance by up to 30%. Ignoring this aspect leads to cross-socket traffic, which saturates the interconnect and degrades overall system responsiveness.

Storage I/O Optimization with NVMe

NVMe drives offer significantly higher IOPS compared to traditional SATA SSDs or HDDs. However, they require specific filesystem and scheduler adjustments to reach their potential.

Filesystem Selection and Mount Options

Choose the ext4 or XFS filesystem for optimal balance between stability and performance. Avoid older filesystems like ext3. When mounting the drive, use the noatime option to prevent unnecessary write operations when files are read.

For databases, consider enabling data=writeback in ext4 mount options. This reduces journaling overhead but increases risk during power loss. Ensure you have battery-backed RAID controllers or UPS systems if choosing this path.

I/O Scheduler Tuning

The Linux I/O scheduler determines how read/write requests are handled. For NVMe devices, the none (or noop) scheduler is often the best choice. NVMe drives handle internal queuing and parallelism efficiently.

Adding an external scheduler layer introduces overhead without providing benefit. Switching to the 'none' scheduler allows the hardware to manage request ordering directly. This change can reduce average latency by 10-15% under heavy load.

Monitor I/O wait times using iostat to verify improvements. High iowait percentages indicate bottlenecks that need addressing through queue depth adjustments or application-level caching.

Network Stack Enhancements for 1Gbps Bandwidth

With 300TB of monthly traffic, network efficiency is critical. The 1Gbps interface must be configured to handle bursty traffic without dropping packets.

TCP/IP Stack Tuning

Adjust kernel parameters in /etc/sysctl.conf to optimize TCP window sizes. Increase net.core.rmem_max and net.core.wmem_max to allow larger buffers. This helps maintain high throughput on high-latency connections.

Enable TCP BBR congestion control algorithm instead of the default Cubic. BBR models the bottleneck bandwidth and round-trip propagation time. It performs better in networks with packet loss or high latency, common in large-scale data transfers.

Interrupt Coalescence and RSS

Configure Network Interface Card (NIC) settings to balance interrupt handling. Use ethtool to enable Receive Side Scaling (RSS). This distributes network interrupts across multiple CPU cores.

Without RSS, a single core handles all network interrupts, creating a bottleneck. Distributing the load ensures that no single CPU core becomes saturated. Adjust interrupt coalescence settings to reduce CPU usage while maintaining low latency.

Industry Context and Broader Implications

Optimizing bare metal servers contrasts sharply with managed cloud services. Cloud providers abstract away hardware details, limiting direct control over kernel parameters.

Enterprises choosing bare metal gain performance predictability. This is crucial for financial trading platforms, real-time analytics, and AI model serving. Unlike shared cloud instances, bare metal eliminates 'noisy neighbor' issues.

The shift toward specialized hardware acceleration continues. While GPUs dominate AI training, CPUs remain vital for data preprocessing and inference. Efficient CPU utilization reduces total cost of ownership (TCO).

What This Means for Developers and Businesses

Developers must design applications with hardware awareness. Code that assumes infinite resources will fail under optimized constraints. Profiling tools should identify NUMA violations and cache misses early in development.

Businesses benefit from reduced infrastructure costs. A well-tuned server handles more concurrent users without scaling horizontally. This delays the need for additional hardware purchases.

Security teams must also adapt. Disabling power-saving features may increase heat output. Data centers need adequate cooling capacity. Monitoring tools should track thermal throttling events to prevent hardware damage.

Looking Ahead: Future Optimization Trends

Future Ubuntu releases will likely integrate AI-driven tuning agents. These tools will automatically adjust kernel parameters based on workload patterns. Manual tuning may become less common as automation improves.

Hardware advancements will continue to push boundaries. Next-generation Xeon processors will feature more cores and integrated accelerators. Software stacks must evolve to utilize these new instructions sets effectively.

Containerization technologies like Kubernetes will play a larger role. Orchestrators will increasingly manage NUMA affinity and huge pages automatically. Developers should focus on writing portable code that leverages these abstractions.

Gogo's Take

• 🔥 Why This Matters: Raw hardware specs are meaningless without proper software configuration. A poorly tuned Xeon Gold server performs worse than a well-tuned consumer-grade CPU. Proper optimization directly translates to lower operational costs and higher user satisfaction by reducing latency.
• ⚠️ Limitations & Risks: Aggressive tuning can destabilize systems. Disabling power management increases energy consumption and heat output. Always test changes in a staging environment before applying them to production. Incorrect NUMA binding can cause severe performance degradation rather than improvement.
• 💡 Actionable Advice: Start by benchmarking your current baseline using sysbench and fio. Apply kernel parameter changes incrementally, testing after each modification. Prioritize NUMA awareness and NVMe scheduler tuning, as these offer the highest return on investment for this specific hardware configuration.