Direct-FUSE: Removing the Middleman for High-Performance FUSE File System Support Yue Zhu*, Teng Wang*, Kathryn Mohror+, Adam Moody+, Kento Sato+, Muhib Khan*, Weikuan Yu* Florida State University* Lawrence Livermore National Laboratory+ Outline • Background & Motivation ·Design ·Performance Evaluation ·Conclusion ROSS’18 S-2 Introduction n High-performance computing (HPC) systems needs efficient file system for supporting large-scale scientific applications Ø Different file systems are used for different kinds of data in a single job Ø Both kernel- and user-level file systems can be used in the applications Ø Due to kernel-level file systems’ development complexity, reliability and portability issues, user-level file systems are more leveraged for particular I/O workloads with special purpose n Filesystem in UserSpacE (FUSE) Ø A software interface for UniX-like computer operating systems Ø It allows non-privileged users to create their own file systems without modifying kernel code Ø User defined file system is run as a separate process in user-space Ø Example: SSHFS, GlusterFS client, FusionFS(BigData’14) ROSS’18 S-3 How does FUSE Work? n Execution path of a function call Ø Send the request to the user-level file system process o App program → VFS → FUSE kernel module → User-level file system process Ø Return the data back to the application program o User-level file system process → FUSE kernel module → VFS → App program Application Program User Level File System User Space Kernel Space Virtual File System (VFS) FUSE Ext4 Page Cache Storage Device ROSS’18 S-4 FUSE File System vs. Native File System FUSE File System Native File System # User-kernel 4 2 Mode Switch # Context 2 0 Switch # Memory 2 1 Copies Application Program User Level File System User Space Kernel Space Virtual File System (VFS) FUSE Ext4 Page Cache Storage Device ROSS’18 S-5 Number of Context Switches & I/O Bandwidth n The complexity added in FUSE file system execution path causes performance degradation in I/O bandwidth Ø tmpfs: a file system that stores data in volatile memory Ø FUSE-tmpfs: a FUSE file system deployed on top of tmpfs Ø dd micro-benchmark and perf system profiling tool are used to gather the I/O bandwidth and the number of context switches Ø Experiment method: continually issue 1000 writes Write Bandwidth # Context Switches Block FUSE-tmpfs tmpfs FUSE- tmpfs Size (KB) (MB/s) (GB/s) tmpfs 4 163 1.3 1012 7 16 372 1.6 1012 7 64 519 1.7 1012 7 128 549 2.0 1012 7 256 569 2.4 2012 7 ROSS’18 S-6 Breakdown of Metadata & Data Latency n The actual file system operations (i.e. metadata or data operations) only occupy a small amount of total execution time Ø Tests are on tmpfs and FUSE-tmpfs Ø Real Operation in metadata operation: the time of conducting operation Ø Data Movement: the actual time of write in a complete write function call Ø Overhead: the cost besides the above two, e.g. the time of context switches 250 11.18% 600 Real Operation Data Movement 38.12% 200 500 Overhead Overhead 150 400 (ns) 300 100 37.86% 200 2.17% 33.7% Latency 50 Latency (ns) 100 34.8% 10.08%15.82% 0 tmpfs FUSE-tmpfs tmpfs FUSE-tmpfs 0 Create Close 1 4 16 64 128 256 Metadata Operations Transfer Sizes (KB) Fig. 1. Time Expense in Metadata Operations Fig. 2. Time Expense in Data Operations ROSS’18 S-7 Existing Solution and Our Approach n How to reduce the overheads from FUSE? Ø Build an independent user-space library to avoid going through kernel (e.g., IndexFS (SC’14), FusionFS) Ø However, this approach cannot support multiple FUSE libraries with distinct file paths and file descriptors n We propose Direct-FUSE to support multiple backend I/O services to an application Ø We adapted libsysio to our purpose in Direct-FUSE o libsysio is developed by Scalability team of Sandia National Lab): « a POSIX-like file I/O, and name space support for remote file systems from an application’s user-level address space. ROSS’18 S-8 Outline ·Background & Motivation • Design ·Performance Evaluation ·Conclusion ROSS’18 S-9 The Overview of Direct-FUSE n Direct-FUSE mainly consists of three components 1. Adapted-libsysio o Intercept file path and file descriptor for backend services identification o Simplify metadata and data execution path in original libsysio 2. lightweight-libfuse (not real libfuse) o Abstract file system operations from backend services to unified APIs 3. Backend services o Provide defined file system operations (e.g., FusionFS) Application Program Direct-FUSE Adapted-libsysio lightweight-libfuse Backend Services FUSE-Ext4 FusionFS Client …. Ext4 FusionFS Server … ROSS’18 S-10 Path and File Descriptor Operations n To facilitate the interception of file system operations for multiple backends, the operations are categorized into two: 1. File path operations i. Intercept prefiX and path (e.g., sshfs:/sshfs/test.txt) and return mount information ii. Look up corresponding inode based on the mount information, and redirect to defined operations 2. File descriptor operations i. Find open-file record based on given file descriptor « Open-file record contains pointers to inode, current stream position, etc ii. Redirect to defined operations based inode info in open-file record ROSS’18 S-11 Requirements for New Backends n Interact with FUSE high-level APIs n Separated as an independent user-space library Ø The library contains the fuse file system operations, initialization function, and also the unmount function Ø If a backend passes some specialized data to the fuse module via fuse_mount(), then the data has to be globalized for later file system operations n Implemented in C/C++ or has to be binary compatible with C/C++ ROSS’18 S-12 Outline ·Background and Challenges ·Design • Performance Evaluation ·Conclusion ROSS’18 S-13 Experimental Methodology n We compare the bandwidth of Direct-FUSE with local FUSE file system and native file system on disk and memory by Iozone Ø Disk o Ext4-fuse: FUSE file system overlying Ext4 o Ext4-direct: Ext4-fuse bypasses the FUSE kernel o Ext4-native: original Ext4 on disk Ø Memory o tmpfs-fuse, tmpfs-direct, and tmpfs-native are similar to the three tests on disk n We also compare the I/O bandwidth of distributed FUSE file system with Direct-FUSE Ø FusionFS: a distributed file system that supports metadata- and write-intensive operations ROSS’18 S-14 Sequential Write Bandwidth n Direct-FUSE achieves comparable bandwidth performance to the native file system Ø Ext4-direct outperforms Ext4-fuse by 16.5% on average Ø tmpfs-direct outperforms tmpfs-fuse at least 2.15x Ext4-fuse Ext4-direct Ext4-native tmpfs-fuse tmpfs-direct tmpfs-native 10000 1000 (MB/s) 100 10 1 Bandwidth 4 16 64 256 1024 Write Transfer Sizes (KB) ROSS’18 S-15 Sequential Read Bandwidth n Similar to the sequential write bandwidth, the read bandwidth of Direct-FUSE is comparable to the native file system Ø Ext4-direct outperforms Ext4-fuse by 2.5% on average Ø tmpfs-direct outperforms tmpfs-fuse at least 2.26X Ext4-fuse Ext4-direct Ext4-native tmpfs-fuse tmpfs-direct tmpfs-native 10000 1000 (MB/s) 100 10 1 Bandwidth 4 16 64 256 1024 Read Transfer Sizes (KB) ROSS’18 S-16 Distributed I/O Bandwidth n Direct-FUSE outperforms FusionFS in write bandwidth and shows comparable read bandwidth Ø Writes benefit more from the FUSE kernel bypassing n Direct-FUSE delivers similar scalability results as the original FusionFS 10000 10000 fusionfs direct-fusionfs fusionfs direct-fusionfs 1000 1000 100 (MB/s) 100 10 10 Bandwidth Bandwidth (MB/s) 1 Bandwidth 1 1 2 4 8 16 1 2 4 8 16 Number of Nodes Number of Nodes Write Read ROSS’18 S-17 Overhead Analysis n The dummy read/write occupies less than 3% of the complete I/O function time in Direct-FUSE, even when the I/O size is very small Ø Dummy write/read: no actual data movement, directly return once reach the backend service Ø Real write/read: the actual Direct-FUSE read and write I/O calls 10000 10000 dummy write real write dummy read real read s) 1000 s) 1000 n n 100 100 Latency ( Latency 10 ( Latency 10 1 1 1B 4B 16B 64B 256B 1KB 1B 4B 16B 64B 256B 1KB Transfer Sizes Transfer Sizes ROSS’18 S-18 Conclusions Ø We have revealed and analyzed the context switches count and time overheads in FUSE metadata and data operations Ø We have designed and implemented Direct-FUSE, which can avoid crossing kernel boundary and support multiple FUSE backends simultaneously Ø Our experimental results indicate that Direct-FUSE achieves significant performance improvement compared to original FUSE file systems ROSS’18 S-19 Sponsors of This Research ROSS’18 S-20.