This study investigates a novel operating system architecture that incorporates the characteristics of next-generation computing accelerators and SCM, aiming to efficiently process intelligent computing tasks on data streams delivered from edge devices at edge-micro data centers. Based on this architecture, we develop an edge-friendly operating system.
| No. | Name | Description |
|---|---|---|
| 1 | bidirectional-readahead | Linux kernel for Bidirectional Readahead |
| 2 | CorePartitioning | A core-partitioning library and kernel module designed to improving I/O performance |
| 3 | CXL-AutoNUMA | Implementation of a dynamic hot-page management technique to support large-scale tenants in EMDC |
| 4 | EMDC_llvm | A utility that provides transparency and dynamic rebinding for heterogeneous accelerators on EMDCs |
| 5 | gpu-faas | A serverless computing platform that minimizes start-up latency for accelerator-based applications by sharing accelerator contexts and thus eliminating context initialization time |
| 6 | inference_profiler | A profiler that can capture the full stack (from the inference layer to the OS kernel) for inference workloads using the NVIDIA Triton Inference Server |
| 7 | KuScale | A scheduler that ensures low latency and provides fine-grained control of GPUs |
| 8 | linux-autozg | A performance optimization technique for ZNS SSDs using dynamic zone group management in F2FS (SYSTOR '24) |
| 9 | mg-lru | A script that analyzes the kernel memory management layer for high-latency I/O requests occurring during read/write operations in low-memory environments (ICCD '23) |
| 10 | pm-aware-scheduler | A dynamic workload relocation technique that takes SCM I/O awareness into account in NUMA environments |
| 11 | pm-opzero | A filesystem that eliminates the data-initialization (zero-out) overhead caused by removing the page cache in DAX |
| 12 | power-aware-trition | “A power management technique incorporating inference energy consumption characteristics in a single-GPU node build on the NVIDIA Triton Inference Server (HPCA '23) |
| 13 | remote-io | A direct access extension model for Remote I/O in the Linux kernel |
| 14 | remote-swap | Code that expands memory and improves application switching performance in mobile systems by leveraging the abundant storage resources of edge servers |
| 15 | severless-rewind | A technique to address security issues arising from function execution in a serverless edge cloud environment while simultaneously optimizing execution performance (USENIX ATC '24) |
| 16 | usbip-COMPRESS | An efficient data transfer implementation based on direct transmit/receive at the kernel layer |