Aerie: Flexible File-System Interfaces to Storage-Class Memory [Eurosys 2014] Operating System Design Yongju Song

Transcription

1 Aerie: Flexible File-System Interfaces to Storage-Class Memory [Eurosys 2014] Operating System Design Yongju Song

2 Outline 1. Storage-Class Memory (SCM) 2. Motivation 3. Design of Aerie 4. File System Features 5. Functionality: PXFS / Optimization: FlatFS 6. Evaluation 7. Conclusion 2

3 Storage-Class Memory (SCM) Phase-Change Memory (PCM), Spin-Transfer-torque MRAM, Flash-backed DRAM Persistent, Shot access time, Byte addressable Storage-Class Memory attached to the CPU Parameter DRAM NAND PCM Density 1x 4x 2x-4x Read Latency 60ns 25μs 200~300ns Write Speed ~1GB/s 2.4MB/s ~100MB/s Byte-addressable Yes No Yes Non-Volatility No Yes Yes [1] Dulloor, Subramanya R., et al. "System software for persistent memory, In Proc. of the European Conference on Computer Systems. ACM,

4 Storage-Class Memory (SCM) Previous work on Kernel-mode File systems Removing some of the kernel layers (e.g., Scheduling, drivers) for reduce access latency to SCM Enabling user-mode access to file data through memory mapping files and user-mode DMA BPFS, SCMFS, PMFS, Quill, Moneta-D à Conventional POSIX interface can limit the performance of SCM Storage-Class Memory no longer requires the OS kernel to provide the file system interface to storage load, store interface of SCM à Doesn t require a kernel driver Direct access from user mode à Can be protected by existing memory-translation H/W Low latency near DRAM à Unnecessary Shared caching 4

5 Storage-Class Memory (SCM) Previous work on Kernel-mode File systems Accessible via loas/stores Removing some of the kernel layers (e.g., Scheduling, drivers) for reduce access latency to SCM Enabling user-mode access to file data through memory mapping files and user-mode DMA BPFS, SCMFS, PMFS, Quill, Moneta-D à Inefficient POSIX interface can limit the benefits Storage-Class Memory no longer requires the OS kernel to provide the file system interface to storage load, store interface of SCM à Doesn t require a kernel driver Direct access from user mode à Can be protected by existing memory-translation H/W Low latency near DRAM à Unnecessary Shared caching 5

6 Motivation POSIX file system interface introduce two sources of overehead Changing modes and cache pollution from entering the kernel Abstract every system resource as a file Quantifying the file abstraction cost Focus on namespace operations (metadata intensive) stat: 25x slower than reading from PCM *in Ext4 on a RAM-disk File system interfaces provides useful features Data organization for easy access Protecting data between applications 6

7 Design of Aerie Main goal Providing applications with flexibility in defining their own FS interface Implementation for high-performance access to SCM 7

8 Design of Aerie Direct Access Direct access through memory to FS data and metadata from user mode FS libraries A Malicious program can corrupt and violate FS invariants As inserting two files with the same name in a directory Decentralized Architecture Untrusted user-mode library (libfs) Virtual memory protection hardware enforces access filesystem data and read-only access to metadata Trusted file system service (TFS) Supports co-operation between distrustful programs Integrity for metadata updates Concurrency control between processes SCM Manager Provides a storage allocation mechanism (Independent of high-level file system organization) 8

9 File System Features Low-level Naming Each file-system storage object is identified by a 64-bit storage object ID 6 bit encode the type of the object / 58 bit encode the virtual memory address High-level Naming collection object provides an associative interface for storing key-value pairs Key à Object ID Indexing memory file (mfile) object type for mapping offsets to extents mfile provides access to a range of bytes starting at a specified offset Can be used to implement data files 9

10 Functionality: PXFS / Optimization: FlatFS PXFS is a POSIX-style interface for compatibility with existing code open/read/write/unlink Files are mfiles with page-sized extents Directories are collections mapping file names to the object IDs A root collection holds the root directory Can access a file concurrent FlatFS shows how to optimize the interface for a specific workload A key-value store interface targeting applications mfiles containing a single extent holding the file contents KV-store concurrency semantics 10

11 Evaluation Generic POSIX performance Does a standard file-system interface based on Aerie perform fast enough to replace kernel file systems? Optimizations Can Aerie provide new opportunities for optimizations exceeding kernel FS performance? Methodology 2.4GHz Intel Xeon E Core Machine equipped with 40GB of DRAM Use a 24GB DRAM memory partition as SCM (add delays using a software spin loop) Workloads Compared to RamFS, Ext3, and Ext4 (Uses the VFS page cache) For application-level workloads, use modified version of Filebench that calls through libfs 11

12 Evaluation PXFS benefits by not calling into the kernel PXFS is faster for open as it benefits from accessing metadata directly without invoking the kernel Latency of common file system operations Average latency per workload operation 12

13 Evaluation Client and Server Scalability Sensitivity to Memory Performance 13

14 Conclusion New storage, called SCM, promise high-speed access to storage directly form user mode Conventional interface is unsuitable for SCM Supporting flexible interfaces, Aerie is proposed Achieve high performance without changes to complex kernel code 14

15 THANK YOU :D Q & A 15