Exploiting Gray-Box Knowledge of Buffer Cache Management Nathan C. Burnett, John Bent, Andrea C. Arpaci-Dusseau, Remzi H. Arpaci-Dusseau University of Wisconsin - Madison Department of Computer Sciences Caching • Buffer cache impacts I/O performance – Cache hits much faster than disk reads Without Cache Knowledge: 2 disk reads With Cache Knowledge 1 disk read OS Buffer Cache Data Blocks 2 Knowledge is Power • Applications can use knowledge of cache state to improve overall performance – Web Server – Database Management Systems • Often no interface for finding cache state – Abstractions hide information 3 Workload + Policy Contents • Cache contents determined by: – Workload – Replacement policy • Algorithmic Mirroring – Observe workload – Simulate cache using policy knowledge – Infer cache contents from simulation model 4 Gaining Knowledge • Application knows workload – Assume application dominates cache • Cache policy is usually hidden – Documentation can be old, vague or incorrect – Source code may not be available • How can we discover cache policy? 5 Policy Discovery • Fingerprinting: automatic discovery of algorithms or policies (e.g. replacement policy, scheduling algorithm) • Dust - Fingerprints buffer cache policies – Correctly identifies many different policies – Requires no kernel modification – Portable across platforms 6 This Talk • Dust – – – – Detecting initial access order (e.g. FIFO) Detecting recency of access (e.g. LRU) Detecting frequency of access (e.g. LFU) Distinguishing clock from other policies • Fingerprints of Real Systems – NetBSD 1.5, Linux 2.2.19, Linux 2.4.14 • Exploiting Gray-Box Knowledge – Cache-Aware Web Server • Conclusions & Future Work 7 Dust • Fingerprints the buffer cache – Determines cache size – Determines cache policy – Determines cache history usage • Manipulate cache in controlled way – open/read/seek/close 8 Replacement Policies • Cache policies often use – access order – recency – frequency • Need access pattern to identify attributes • Explore in simulation – Well controlled environment – Variety of policies – Known implementations 9 Dust I. Move cache to known state a. b. c. Sets initial access order Sets access recency Sets frequency II. Cause part of test data to be evicted III. Sample data to determine cache state • Read a block and time it Repeat for confidence 10 Setting Initial Access Order Test Region Eviction Region for ( 0 test_region_size/read_size) { read(read_size); } 11 FIFO Priority Newer Pages Older Pages FIFO gives latter part of file priority 12 Detecting FIFO Out of Cache In Cache • FIFO evicts the first half of test region 13 Setting Recency Test Region Left Pointer Eviction Region Right Pointer do_sequential_scan(); left = 0; right = test_region_size/2; for ( 0 test_region_size/read_size){ seek(left); read(read_size); seek(right); read(read_size); right+=read_size; left+= read_size; } 14 LRU Priority LRU gives priority to 2nd and 4th quarters of test region 15 Detecting LRU • LRU evicts 1st and 3rd quarters of test region 16 Setting Frequency Test Region 2 3 4 5 6 6 5 4 3 2 Left Pointer Eviction Region 7 Right Pointer do_sequential_scan(); left = 0; right = test_region_size/2; left_count = 1; right_count = 5; for ( 0 test_region_size/read_size) for (0 left_count) seek(left); read(read_size); for (0 right_count) seek(right); read(read_size); right+=read_size; left+= read_size; 17 right_count++; left_count--; LFU Priority LFU gives priority to center of test region 18 Detecting LFU • LFU evicts outermost stripes • Two stripes partially evicted 19 The Clock Algorithm • • • • Used in place of LRU Ref. bit set on reference Ref. bit cleared as hand passes Hand replaces a page with a ref. bit that’s already clear • On eviction, hand searches for a clear ref. bit Page Frame Reference bit 20 Detecting Clock Replacement • Two pieces of initial state – Hand Position – Reference Bits • Hand position is irrelevant – circular queue • Dust must control for reference bits – Reference bits affect order of replacement 21 Detecting Clock Replacement • Uniform reference bits • Random reference bits 22 Clock - Reference Bits Matter • • • • Two fingerprints for Clock Ability to produce both will imply Clock Need a way to selectively set reference bits Dust manipulates reference bits – To set bits, reference page – To clear all bits, cause hand to sweep • Details in paper 23 Dust Summary • Determines cache size (needed to control eviction) • Differentiates policies based on – access order – recency – frequency • Identifies many common policies – FIFO, LRU, LFU, Clock, Segmented FIFO, Random • Identifies history-based policies – LRU-2, 2-Queue 24 This Talk • Dust – Detecting initial access order (e.g. FIFO) – Detecting recency of access (e.g. LRU) – Detecting frequency of access (e.g. LFU) – Distinguishing clock from other policies • Fingerprints of Real Systems – NetBSD 1.5, Linux 2.2.19, Linux 2.4.14 • Exploiting Gray-Box Knowledge – Cache-Aware Web Server • Conclusions & Future Work 25 Fingerprinting Real Systems • Issues: – Data is noisy – Policies usually more complex – Buffer Cache/VM Integration • Cache size might be changing • Platform: – Dual 550 MHz P-III Xeon, 1GB RAM, Ultra2 SCSI 10000RPM Disks 26 F I F O NetBSD 1.5 L R U L F U • Increased variance due to storage hierarchy 27 F I F O NetBSD 1.5 L R U L F U • Four distinct regions of eviction/retention 28 F I F O NetBSD 1.5 L R U L F U • Trying to clear reference bits makes no difference • Conclusion: LRU 29 F I F O Linux 2.2.19 L R U L F U • Very noisy but looks like LRU • Conclusion: LRU or Clock 30 F I F O Linux 2.2.19 L R U L F U • Clearing Reference bits changes fingerprint • Conclusion: Clock 31 F I F O Linux 2.4.14 L R U L F U • Low recency areas are evicted • Low frequency areas also evicted • Conclusion: LRU with page aging 32 This Talk • Dust – Detecting initial access order (e.g. FIFO) – Detecting recency of access (e.g. LRU) – Detecting frequency of access (e.g. LFU) – Distinguishing clock from other policies • Fingerprints of Real Systems – NetBSD 1.5, Linux 2.2.19, Linux 2.4.14 • Exploiting Gray-Box Knowledge – Cache-Aware Web Server • Conclusions & Future Work 33 Algorithmic Mirroring • Model Cache Contents – Observe inputs to cache (reads) – Use knowledge of cache policy to simulate cache • Use model to make application-level decisions 34 NeST • • • • NeST - Network Storage Technology Software based storage appliance Supports HTTP, NFS, FTP, GridFTP, Chirp Allows configurable number of requests to be serviced concurrently • Scheduling Policy: FIFO 35 Cache-Aware NeST • Takes policy & size discovered by Dust • Maintains algorithmic mirror of buffer cache – Updates mirror on each request – No double buffering – May not be a perfect mirror • Scheduling Policy: In-Cache-First – Reduce latency by approximating SJF – Improve throughput by reducing disk reads 36 Performance 144 clients randomly requesting 200, 1MB files Server: P-III Xeon, 128MB Clients: 4 X P-III Xeon, 1GB Gigabit Ethernet Linux 2.2.19 • Improvement in response time • Robust to inaccuracies in cache estimate 37 Summary • Fingerprinting – Discovers OS algorithms and policies • Dust – Portable, user-level cache policy fingerprinting – Identifies FIFO, LRU, LFU, Clock, Random, 2Q, LRU-2 – Fingerprinted Linux 2.2 & 2.4, Solaris 2.7, NetBSD 1.5 & HP-UX 11.20 • Algorithmic Mirroring – Keep track of kernel state in user-space – Use this information to improve performance • Cache-Aware NeST – Uses mirroring to improved HTTP performance 38 Future Work • On-line, adaptive detection of cache policy • Policy manipulation • Make other applications cache aware – Databases – File servers (ftp, NFS, etc.) • Fingerprint other OS components – CPU scheduler – filesystem layout 39 Questions?? • Gray-Box Systems – http://www.cs.wisc.edu/graybox/ • Wisconsin Network Disks – http://www.cs.wisc.edu/wind/ • NeST – http://www.cs.wisc.edu/condor/nest/ 40 F I F O Solaris 2.7 L R U L F U 41 F I F O HP-UX 11.20 (IPF) L R U L F U • Low recency areas are evicted • Low frequency areas also evicted • Conclusion: LRU with page aging 42 Related Work • Gray-Box (Arpaci-Dusseau) – Cache content detector • Connection Scheduling (Crovella, et. al.) • TBIT (Padhye & Floyd) 43 Clock - Uniform Reference Bits File Buffer Cache before test scan Buffer Cache after test scan, before eviction scan • After initial scan, cache state does not change • First half of test region is evicted 44 Clock - Random Reference Bits File Buffer Cache before test scan Buffer Cache after test scan, before eviction scan • Initial Sequential Scan • Test scan does not change cache state 45 Manipulating Reference Bits Buffer Cache after touching all resident data Buffer Cache after an additional small read • Setting bits is easy • Clear bits by causing hand to do a circuit 46