transparent huge pages ● GC Related ■ case study: GC sweep caused latency jitter ■ case study: Lock and NUMA aware Agenda Part I - Latency in scheduler ● The client consists of a goroutine and a channel ○ Go GC ● Unexpected latency jitter caused by GC may still exist Conclusion Case study: Lock and NUMA aware ● TiDB server CPU usage is around 60% ● Networking and IO are both OK ● Increase the benchmark columns ● The deployment does not consider NUMA ○ CPU usage is higher when binding CPUs to NUMA nodes Reproduce ● Context Switch, Minor Page Fault, and NUMA Hit Miss are related Reproduce Reproduce

NIC GPU GPU GPU GPU Xe LinkMulti-GPU Systems - NUMA regions: - 4+ GPUs - 2+ CPUs CPU NIC GPU GPU GPU GPU Xe LinkMulti-GPU Systems - NUMA regions: - 4+ GPUs - 2+ CPUs - Systems becoming data structures - Support distributed algorithms - Achieve high performance for both multi-GPU, NUMA, and multi-node execution float dot_product(vector& x, vector& y)

16 bytes, and for other data types ● gcc / clang -mcx16 – avoid calling the atomic library ● Numa consideration: – entries array on reader side 19 CppCon 2024 | Erez Strauss | Lockfree, MPMC Queue

Protobuf110 FastQueue – Going even further • Bulk Writing • Duplicate Queue Header on “remote” NUMA nodeMeasurements in low-latency trading systems 115Measurements in low-latency trading systems

属于顺序写，目前我们并没有单独对他进行配置，建议 SSD，RAID 如果允许的话，最好是 RAID 10，RAID 卡 cache、操作系统 I/O 调度目前没有针对性的最佳实践，Linux 7 以上默认配置即可，NUMA 没有特别建议， NUMA 内存分配策略可以尝试使用 interleave = all ，文件系统建议 ext4。 一般来说开启 sync-log 性能损耗大概 30% 左右，官方有个 sync-log innodb_flush_log_trx_commit 参数，来保证提交 数据不丢失？ 3.4.14 对 WAL 存储有什么推荐的硬件配置，例如 SSD，RAID 级别，RAID 卡 cache 策略，NUMA 设置 ,文件系统选择，操作系统的 IO 调度策略等？ 3.4.15 在最严格 sync-log = ture 的数据可用模式下，写入性能如何？ 3.4.16 是否可以利用上层的 Raft + 多副本，达到完全的数据可靠，单机存储

machines into VirtualBox. With its unique ACPI power status support, VirtualBox can even report to ACPI-aware guest operating systems the power status of the host. For mobile systems running on battery, the guest provide the modern PC architecture expected by OS X without requiring any “hacks”. You should be aware of a number of important issues before attempting to install a Mac OS X guest: 1. Mac OS X is commercial provide vmname.pxe. 6.3.3 NAT limitations There are four limitations of NAT mode which users should be aware of: ICMP protocol limitations: Some frequently used network debugging tools (e.g. ping or tracerouting)

machines into VirtualBox. With its unique ACPI power status support, VirtualBox can even report to ACPI-aware guest operating systems the power status of the host. For mobile systems running on battery, the guest provide the modern PC architecture expected by OS X without requiring any “hacks”. You should be aware of a number of important issues before attempting to install a Mac OS X guest: 1. Mac OS X is commercial provide vmname.pxe. 6.3.3 NAT limitations There are four limitations of NAT mode which users should be aware of: ICMP protocol limitations: Some frequently used network debugging tools (e.g. ping or tracerouting)

accept a "policy-aware scheduler." ● This scheduler combines an execution policy ("how") and a scheduler ("where"). ● The execute_on function facilitates the creation of such policy-aware schedulers. Design Design Goals: ● Extending C++ parallel algorithms with policy-aware schedulers. ● Allowing customization for different execution contexts. ● ● Preserve core semantics of algorithms and policies ● Cover Concepts 1. policy_aware_scheduler 2. execute_on 3. Customizable functions ● ● The execution_policy concept defines the requirements for execution policies. ● The policy_aware_scheduler concept represents

40 m / s^2 10 m 20 m / s 30 km / h 31.4Unit-aware inverses https://aurora-opensource.github.io/au/main/reference/math/#inverse-functions 32Unit-aware inverses https://aurora-opensource.github.io/ 1Unit-aware inverses https://aurora-opensource.github.io/au/main/reference/math/#inverse-functions f = 400 Hz T = 1/f = 0.0025 s = 2 500 μs f = 400 Hz T = 1/f = 0.0025 s = 2 500 μs 32.2Unit-aware inverses 0.0025 s = 2 500 μs f = 400 Hz T = 1/f = 0.0025 s = 2 500 μs Solve for (units of) 1... 32.3Unit-aware inverses https://aurora-opensource.github.io/au/main/reference/math/#inverse-functions f = 400 Hz

motivation behind Taro • Learn to use the Taro C++ programming model • Dive into the Taro’s coroutine-aware scheduling algorithm • Evaluate Taro on microbenchmarks and a real-world application • Conclusion motivation behind Taro • Learn to use the Taro C++ programming model • Dive into the Taro’s coroutine-aware scheduling algorithm • Evaluate Taro on microbenchmarks and a real-world application • Conclusion motivation behind Taro • Learn to use the Taro C++ programming model • Dive into the Taro’s coroutine-aware scheduling algorithm • Evaluate Taro on microbenchmarks and a real-world application • Conclusion