Compal Electronics Inc. - BIOS and OS Tuning Descriptions for SPEC CPU2017 Submissions

Operating System Tuning Parameters

OS Tuning

ulimit:

is a command used to set or check user limits on system resources such as memory, CPU, and the number of open files. Below are common usages of ulimit:

• ulimit -s: Sets the stack size limit.
• ulimit -l: Sets the maximum amount of memory that can be locked into RAM.
• ulimit -n: Limits the maximum number of open file descriptors.
• ulimit -u: Limits the maximum number of user processes that can be created.

Performance Governors (Linux):

Performance governors are part of Linux's CPU frequency scaling mechanisms, used to determine how the CPU frequency should be managed. Simply put, they control "how fast the CPU should run under different conditions." Common CPU governors include:

• cpupower frequency-set -g performance: Keeps the CPU running at the highest frequency for maximum performance (but consumes more power).
• cpupower frequency-set -g powersave: Keeps the CPU running at the lowest frequency to save power at the cost of performance.
• cpupower frequency-set -g ondemand: Dynamically switches between high and low frequencies based on system load.
• cpupower frequency-set -g conservative: Similar to ondemand, but changes frequency more gradually for smoother transitions.
• cpupower frequency-set -g schedutil: Adjusts frequency based on information from the task scheduler, balancing performance and power saving.

--governor, -g:

When set to performance, the CPU will always operate at its maximum frequency to deliver the highest computing performance. This will improve overall system performance.

Many companies execute the following command when conducting system performance testing to ensure that the CPU operates at its maximum frequency:

• cpupower frequency-set -g performance

drop_caches:

To clear the Linux filesystem cache during testing or prior to benchmarking, the following command is used:

• sync; sysctl -w vm.drop_caches=3: Forces a filesystem sync to flush dirty pages to disk, then instructs the Linux kernel to free pagecache, dentries, and inodes. This helps minimize the impact of cached data on benchmark results by simulating a clean memory state before workload execution.

dirty_ratio:

When dirty pages exceed this ratio, the kernel starts writing them to disk aggressively:

• Available Values:0–100 (percent)
• Lower value (e.g., 5–10%): Write dirty data earlier, reduce I/O bursts
• Higher value: Allow larger in-memory buffers before flushing

swappiness:

It decides how much the kernel prefers swapping vs. keeping data in RAM.:

• Available Values:0–100
• 0: Avoid swapping as much as possible
• 1: Almost disable swap, only swap when absolutely necessary
• 10: Very swap-averse (common in performance tuning)
• 60: Default on most Linux systems
• 100: Swap aggressively, keep more RAM available for file cache

reclaim_mode:

It forces memory allocations to stay within the local NUMA node when possible.:

• Available Values(bitmask):0,1,2,4
• 0: Disable zone reclaim. The system can freely allocate memory from remote NUMA nodes.
• 1: Reclaim by cleaning dirty pages (writeback).
• 2: Reclaim by swapping anonymous pages.
• 4: Reclaim by unmapped page stealing (take clean pages).

/sys/kernel/mm/transparent_hugepage/enabled:

This setting controls whether Transparent Huge Pages are used automatically by the kernel when allocating memory. It determines if the system should attempt to use huge pages for anonymous memory (e.g., heap, stack) without requiring explicit application support:

• always:The kernel will always try to use THP whenever possible.
• madvise:THP is only used when applications explicitly request it via madvise(MADV_HUGEPAGE).
• never:THP is completely disabled. The kernel will not use huge pages automatically.

/sys/kernel/mm/transparent_hugepage/defrag:

This controls how aggressively the kernel should defragment memory to create huge pages. Defragmentation may cause latency spikes because the kernel must compact memory.:

• always:Always attempt memory compaction to create huge pages, even if it may cause delays.
• madvise:Only perform defragmentation for memory regions marked with madvise(MADV_HUGEPAGE).
• defer:Do not block the process; defer huge page compaction work to background threads.
• defer+madvise:Defer compaction but only for memory explicitly marked via madvise.
• never:Do not attempt memory compaction for huge pages.

Firmware / BIOS / Microcode Settings

SMT Control (Default = Auto):

SMT Control is a setting that enables or disables Simultaneous Multithreading (SMT), allowing each CPU core to execute one or more threads concurrently to improve multitasking performance or ensure thread isolation. Values for this BIOS option can be:

• Auto
• Enabled
• Disabled

Power Profile Selection (Default = Efficiency Mode):

This setting controls how the system balances power efficiency and performance across CPU, memory, and I/O subsystems. Values for this BIOS option can be:

• High Performance Mode: Maximizes CPU frequency and performance at the cost of increased power consumption.
• Efficiency Mode: Reduces power usage by optimizing for lower energy consumption with acceptable performance.
• Maximum IO Performance Mode: Optimizes system settings to prioritize I/O throughput for data-intensive operations.
• Balanced Memory Performance Mode: Balances memory bandwidth and latency for applications sensitive to memory performance.
• Balanced Core Performance Mode: Optimizes CPU core performance while maintaining moderate power efficiency.
• Balanced Core Memory Performance Mode: Provides a balanced configuration between CPU core speed and memory throughput for general workloads.

SVM Mode (Default = Enable):

SVM (Secure Virtual Machine) Mode is a BIOS setting that enables or disables hardware-assisted virtualization on AMD processors. When enabled, it allows the use of virtualization technologies such as AMD-V, which are required by hypervisors (e.g., VMware, Hyper-V, KVM) to run virtual machines with hardware-level isolation and improved performance. Values for this BIOS option can be:

• Enabled
• Disabled

SR-IOV Support (Default = Enabled):

SR-IOV (Single Root I/O Virtualization) is a hardware-assisted virtualization technology that allows a single physical PCIe device (such as a network interface card) to present multiple virtual functions (VFs) to the operating system or hypervisor. This enables more efficient and direct access to hardware for virtual machines, reducing I/O overhead and improving performance in virtualized environments. Values for this BIOS option can be:

• Enabled
• Disabled

TDP Control (Default = Auto):

TDP Control determines how the processor’s Thermal Design Power (TDP) is managed — either automatically by the system or manually by user-defined limits. This setting affects CPU power consumption and thermal behavior. Values for this BIOS option can be:

• Auto
• Manual

TDP (Default = 0):

TDP (Thermal Design Power) sets a power consumption target for the CPU in watts, helping manage thermal output and power limits during operation — especially relevant when TDP Control is set to Manual.

• 0: Uses the processor's default TDP value as defined by AMD (no user-defined override).

PPT Control (Default = Auto):

PPT Control (Package Power Tracking Control) determines whether the maximum allowable CPU package power (PPT limit) is automatically set by the system or manually defined by the user to control CPU power usage. Values for this BIOS option can be:

• Auto
• Manual

PPT (Default = 0):

PPT defines the upper limit of total power consumption (in watts) for the CPU package, including cores, cache, and SoC components, to ensure thermal and electrical safety.

• 0: A value of 0 typically means no user-defined PPT limit, and the system will rely on default platform or processor limits unless overridden elsewhere.

ACPI CST C2 Latency (Default = 100):

ACPI CST C2 Latency defines the response time (in microseconds) for the processor to exit the C2 low-power state and return to full operation. This setting influences how quickly the CPU can resume tasks after being in power-saving mode.

• 100: A latency value of 100 µs means the processor will take approximately 100 microseconds to exit C2 state and resume activity, balancing power savings with wake-up responsiveness.

NUMA Nodes Per Socket (Default = Auto):

NUMA Nodes Per Socket (NPS) determines how many NUMA (Non-Uniform Memory Access) domains are created per CPU socket, impacting memory locality, bandwidth, and latency for multi-threaded workloads. Values for this BIOS option can be:

• Auto: Automatically selects the NUMA configuration based on system firmware or OS recommendations.
• NPS0: Combines all cores and memory of a socket into a single NUMA node to reduce cross-node communication.
• NPS1: Divides the socket into 1 NUMA node per socket (default for many applications).
• NPS2: Splits the socket into 2 NUMA nodes, improving locality for memory-intensive tasks.
• NPS4: Splits the socket into 4 NUMA nodes for maximum memory bandwidth and minimal latency per region, ideal for NUMA-aware applications.

DRAM Scrub Time (Default = 24hr):

DRAM Scrub Time defines the periodic interval for background memory error correction (memory scrubbing), which helps detect and repair soft errors (bit flips) in DRAM to improve system reliability. Values for this BIOS option can be:

• 48hr: Performs DRAM error scrubbing once every 48 hours, minimizing overhead with slower error correction.
• 24hr: Performs DRAM scrubbing every 24 hours (default for balanced protection and overhead).
• 16hr: Triggers memory scrubbing every 16 hours for more frequent ECC correction.
• 12hr: Scrubs DRAM every 12 hours for improved error mitigation.
• 8hr: Increases ECC scanning frequency to once every 8 hours.
• 6hr: Shortens scrub interval to 6 hours to catch errors more proactively.
• 4hr: Runs scrubbing every 4 hours, enhancing protection for high-availability systems.
• 1hr: Aggressively scrubs DRAM every hour, maximizing error correction but with higher CPU/memory overhead.
• Disabled: Disables DRAM background scrubbing entirely, reducing overhead but increasing risk of uncorrected soft errors.

L1 Stride Prefetcher (Default = Auto):

L1 Stride Prefetcher is a processor feature that attempts to pre-load data into the L1 cache by predicting memory access patterns with regular strides, helping improve performance by reducing cache miss latency. Values for this BIOS option can be:

• Auto
• Enabled
• Disabled

APBDIS (Default = 0):

APBDIS (Application Power Brake Disable) is a BIOS setting that controls whether the CPU’s internal power throttling feature (Application Power Brake, APB) is enabled or disabled. APB dynamically reduces performance under certain conditions to meet power or thermal constraints. Values for this BIOS option can be:

• 0: APB is enabled, allowing the processor to throttle performance to reduce power/thermal load.
• 1: APB is disabled, preventing the processor from throttling performance due to APB triggers.
• Auto: The system automatically decides whether to enable or disable APB based on platform policies or workload behavior.

DF Cstates (Default = Auto)

Much like CPU cores, the Infinity Fabric can go into lower power states while idle. However, there will be a delay changing back to full-power mode causing some latency jitter. In a low latency workload, or one with bursty I/O, one could disable this feature to achieve more performance with the tradeoff of higher power consumption:

• Auto
• Enabled
• Disabled

Global C-state Control (Default = Auto)

dictates a CPU's ability to enter deep power-saving states when idle. Enabling it allows the CPU to use its most energy-efficient, deeper C-states (like C6), which saves power and lowers idle temperatures but can add slight latency when waking up. Disabling it forces the CPU to use shallower states (like C1/C1E), increasing power consumption and idle temperatures but may reduce latency for tasks requiring quick wake-up, which can sometimes prevent system stutters or instability in specific scenarios, particularly with certain CPUs

• Auto
• Enabled
• Disabled

Last updated Aug 8, 2025.