Understanding Manycore Scalability of File Systems. Changwoo Min, Sanidhya Kashyap, Steffen Maass Woonhak Kang, and Taesoo Kim

Transcription

1 Understanding Manycore Scalability of File Systems Changwoo Min, Sanidhya Kashyap, Steffen Maass Woonhak Kang, and Taesoo Kim

2 Application must parallelize I/O operations Death of single core CPU scaling CPU clock frequency: ~. GHz # of physical cores: up to (Xeon E7 v) From mechanical HDD to flash SSD IOPS of a commodity SSD: 9K Non-volatile memory (e.g., D XPoint):,x But file systems become a scalability bottleneck

3 Problem: Lack of understanding in internal scalability behavior Exim mail server on AMDISK k messages/sec k btrfs FFS ext XFS k Embarrassingly parallel application!. Saturated k k k k k. Collapsed 7 Intel -core machine: -socket, -core Xeon E7-7 AM: GB, TB SSD, 7 PM HDD. Never scale

4 Even in slower storage medium file system becomes a bottleneck Exim server at cores k AMDISK SSD HDD messages/sec k k k k k k btrfs ext FFS XFS

5 Outline Background FxMark design A file system benchmark suite for manycore scalability Analysis of five Linux file systems Pilot solution elated work Summary

6 esearch questions What file system operations are not scalable? Why they are not scalable? Is it the problem of implementation or design?

7 Technical challenges Applications are usually stuck with a few bottlenecks cannot see the next level of bottlenecks before resolving them difficult to understand overall scalability behavior How to systematically stress file systems to understand scalability behavior 7

8 FxMark: evaluate & analyze manycore scalability of file systems FxMark: File systems: tmpfs Memory FS ext XFS J/NJ Journaling FS Storage medium: # core: applications 9 micro-benchmarks btrfs FFS CoW FS Log FS SSD,,,,,,,,, 7,

9 FxMark: evaluate & analyze manycore scalability of file systems FxMark: File systems: tmpfs Memory FS ext XFS >,7 J/NJ Journaling FS Storage medium: # core: applications 9 micro-benchmarks btrfs FFS CoW FS Log FS SSD,,,,,,,,, 7, 9

10 Microbenchmark: unveil hidden scalability bottlenecks Data block read Low Sharing Level Medium File High Block Process Operation Legend:

11 Stress different components with various sharing levels

12 Evaluation Data block read Linear scalability Low: File systems: Legend btrfs ext extnj FFS tmpfs XFS Storage medium: 7

13 Outline Background FxMark design Analysis of five Linux file systems What are scalability bottlenecks? Pilot solution elated work Summary

14 Summary of results: file systems are not scalable DBM DWSL MPL 7 MPM 7... MPH MWCM k. k messages/sec k k k k k DBM:O_DIECT btrfs ext extnj FFS tmpfs XFS 7. Legend... DBENCH ocksdb 7 MWCL. DBL:O_DIECT.. MWM. k. k. k.. Exim.. MWL. k. DWOM:O_DIECT DWOL:O_DIECT MWUM MWUL. GB/sec ops/sec 7.. MDM MDL DWTL DWAL 7 DWOM DWOL DBH DBL 7

15 Summary of results: file systems are not scalable DBM DWSL MPL 7 MPM 7... MPH MWCM k. k messages/sec k k k k k DBM:O_DIECT btrfs ext extnj FFS tmpfs XFS 7. Legend... DBENCH ocksdb 7 MWCL. DBL:O_DIECT.. MWM. k. k. k.. Exim.. MWL. k. DWOM:O_DIECT DWOL:O_DIECT MWUM MWUL. GB/sec ops/sec 7.. MDM MDL DWTL DWAL 7 DWOM DWOL DBH DBL 7

16 Summary of results: file systems are not scalable DBM DWSL MPL 7 MPM 7... MPH MWCM k. k messages/sec k k k k k DBM:O_DIECT btrfs ext extnj FFS tmpfs XFS 7. Legend... DBENCH ocksdb 7 MWCL. DBL:O_DIECT.. MWM. k. k. k.. Exim.. MWL. k. DWOM:O_DIECT DWOL:O_DIECT MWUM MWUL. GB/sec ops/sec 7.. MDM MDL DWTL DWAL 7 DWOM DWOL DBH DBL 7

17 Summary of results: file systems are not scalable DBM DWSL MPL 7 MPM 7... MPH MWCM k. k messages/sec k k k k k DBM:O_DIECT btrfs ext extnj FFS tmpfs XFS 7. Legend... DBENCH ocksdb 7 MWCL. DBL:O_DIECT.. MWM. k. k. k.. Exim.. MWL. k. DWOM:O_DIECT DWOL:O_DIECT MWUM MWUL. GB/sec ops/sec 7.. MDM MDL DWTL DWAL 7 DWOM DWOL DBH DBL 7

18 Data block read DBL All file systems linearly scale Low: 7 DBM XFS shows performance collapse Medium: XFS 7 DBH All file systems show performance collapse 9 7 High: 7

19 Page cache is maintained for efficient access of file data OS Kernel. look up a page cache. read a file block Page cache. copy page. cache miss. read a page from disk Disk 9

20 Page cache hit OS Kernel. look up a page cache. read a file block Page cache. copy page. cache hit Disk

21 Page cache can be evicted to secure free memory OS Kernel Page cache Disk

22 only when not being accessed OS Kernel. read a file block eference counting is used to track # of accessing tasks Page cache. copy page access_a_page(...) { atomic_inc(&page->_count);... atomic_dec(&page->_count); } Disk

23 eference counting becomes a scalability bottleneck CPI DBH (cycles-per-instruction) access_a_page(...) { atomic_inc(&page->_count);... atomic_dec(&page->_count); } CPI (cycles-per-instruction) 7

24 eference counting becomes a scalability bottleneck CPI DBH (cycles-per-instruction) access_a_page(...) { atomic_inc(&page->_count);... atomic_dec(&page->_count); } CPI (cycles-per-instruction) High contention on a page reference counter Huge memory stall 7 Many more: directory entry cache, XFS inode, etc

25 Lessons learned High locality can cause performance collapse Cache hit should be scalable When the cache hit is dominant, the scalability of cache hit does matter.

26 Data block overwrite DWOL W Ext, FFS, and btrfs show performance collapse Low: W ext FFS btrfs 7 DWOM All file systems degrade gradually... Medium: W W

27 Btrfs is a copy-on-write (CoW) file system Directs a write to a block to a new copy of the block Never overwrites the block in place Maintain multiple versions of a file system image W Time T Time T Time T+

28 CoW triggers disk block allocation for every write W Time T Time T+ Block Allocation Block Allocation Disk block allocation becomes a bottleneck Ext journaling, FFS checkpointing

29 Lessons learned Overwriting could be as expensive as appending Critical at log-structured FS (FFS) and CoW FS (btrfs) Consistency guarantee mechanisms should be scalable Scalable journaling Scalable CoW index structure Parallel log-structured writing 9

30 Data block overwrite DWOL W Ext, FFS, and btrfs show performance collapse Low: W ext FFS btrfs 7 DWOM All file systems degrade gradually... Medium: W W

31 Entire file is locked regardless of update range All tested file systems hold an inode mutex for write operations ange-based locking is not implemented ***_file_write_iter(...) { mutex_lock(&inode->i_mutex);... mutex_unlock(&inode->i_mutex); }

32 Lessons learned A file cannot be concurrently updated Critical for VM and DBMS, which manage large files Need to consider techniques used in parallel file systems E.g., range-based locking

33 Summary of findings High locality can cause performance collapse Overwriting could be as expensive as appending A file cannot be concurrently updated All directory operations are sequential enaming is system-wide sequential Metadata changes are not scalable Non-scalability often means wasting CPU cycles Scalability is not portable See er p a p our

34 Summary of findings Many themcan arecause unexpected and collapse counter-intuitive High of locality performance Contention at file system level Overwriting could be as expensive as appending to maintain data dependencies A file cannot be concurrently updated All directory operations are sequential enaming is system-wide sequential Metadata changes are not scalable Non-scalability often means wasting CPU cycles Scalability is not portable See er p a p our

35 Outline Background FxMark design Analysis of five Linux file systems Pilot solution If we remove contentions in a file system, is such file system scalable? elated work Summary

36 ocksdb on a -partitioned AMDISK scales better A single-partitioned AMDISK tmpfs 7 7 ops/sec ops/sec A -partitioned AMDISK 7.x ** Tested workload: DB_BENCH overwrite ** btrfs ext FFS tmpfs XFS 7

37 ocksdb on a -partitioned AMDISK scales better A single-partitioned AMDISK tmpfs 7 7 ops/sec ops/sec A -partitioned AMDISK 7.x btrfs ext FFS tmpfs XFS 7 educed contention on file systems ** Tested workload: DB_BENCH overwrite helps improving performance and**scalability 7

38 But partitioning makes performance worse on HDD ops/sec A single-partitioned HDD FFS A -partitioned HDD btrfs ext FFS XFS.7x ** Tested workload: DB_BENCH overwrite **

39 But partitioning makes performance worse on HDD ops/sec A single-partitioned HDD FFS A -partitioned HDD btrfs ext FFS XFS.7x But reduced spatial locality degrades performance Medium-specific (e.g.,**spatial locality) ** Testedcharacteristics workload: DB_BENCH overwrite should be considered 9

40 elated work Scaling operating systems Mostly use memory file system to opt out the effect of I/O operations Scaling file systems Scalable file system journaling ScaleFS [MIT:MSThesis'] SpanFS [ATC'] Parallel log-structured writing on NVAM NOVA [FAST']

41 Summary Comprehensive analysis of manycore scalability of five widely-used file systems using FxMark Manycore scalability should be of utmost importance in file system design New challenges in scalable file system design Minimizing contention, scalable consistency guarantee, spatial locality, etc. FxMark is open source