Deconstructing Commodity Storage Clusters Haryadi S. Gunawi, Nitin Agrawal, Andrea C. Arpaci-Dusseau, Remzi H. Arpaci-Dusseau Univ. of Wisconsin - Madison Jiri Schindler Corporation Storage system Storage system – Important components of large-scale systems – Multi-billion dollar industry Often comprised of high-end storage servers – A big box with lots of disks inside The simple question – How does storage server work? – Simple but hard – closed storage subsystem design 2 Why need to know? Better modeling – How system behaves under different workload – Example in storage industry: capacity model for capacity planning – Model is limited if the information is limited Product validation – Validate what product specs say – Performance numbers cannot confirm Critical evaluation of design and implementation choices – Control what is occurring inside 3 Traditionally black box Highly customized and proprietary hardware and OS – Hitachi Lightning, NetApp Filers, EMC Symmetrix – EMC Symmetrix: disk/cache manager, proprietary OS Internal information is hidden behind standard interfaces Storage System Client Acks ? 4 Modern graybox storage system Cluster of commodity PCs running commodity OS – Google FS cluster, HP FAB, EMC Centera Advantages of commodity storage clusters – Direct internal observation – visible probe points – Leverage existing standardized tools Storage System Update DB Client 5 Intra-box Techniques Two “Intra-box” techniques – Observation – System perturbation Two components of analysis – Deduce structure of main communication protocol • Object Read and Write protocol – Internal policy decisions • Caching, prefetching, write buffering, load balancing, etc. 6 Goal and EMC Author Objectives – Feasibility of deconstructing commodity storage clusters, no source code – Results achieved without EMC assistance EMC Author – Evaluate correctness of our findings – Give insights behind their design decisions 7 Outline Introduction EMC Centera Overview – Intra-box tools Deducing Protocol – Observation and Delay Pertubation Inferring Policies – System Perturbation Conclusion 8 Centera Topology Storage Nodes Access Nodes Client SN 1 SN 2 Client WAN AN 1 AN 2 SN 3 LAN SN 4 SN 5 SN 6 9 Commodity OS Client Client SDK TCP Access Node Storage Node Centera Software Centera Software Linux Linux TCP/UDP Reiserfs IDE driver WAN TCP/UDP Reiserfs IDE driver LAN 10 Probe Points – Observation Client Access Node Storage Node Client SDK Centera SW. Centera Software TCP TCP/UDP TCP/UDP Reiserfs tcpdump tcpdump tcpdump Pseudo Dev. Driver IDE drives Internal probe points – Trace traffic using standardized tools • • tcpdump: trace network traffic Pseudo Device Driver: trace disk traffic 11 Probe Points – Perturbation Storage Node User-level Process Client Access Node Client SDK Centera SW TCP TCP/UDP TCP/UDP Reiserfs Mod. NistNet Mod. NistNet Mod. NistNet + Delay tcpdump tcpdump tcpdump Pseudo Dev. Centera Software Add CPU Load: while(1) {..} Add Disk Load: cp fX fY IDE drives Perturbing system at probe points – Modified NistNet: delay particular messages – Pseudo Dev. Driver: delay disk I/O traffic – Additional Load • • CPU Load: High priority while loop Disk Load: File copy 12 Outline Introduction EMC Centera Overview Deducing Protocol – Observation and Delay Perturbation Inferring Policies – System Perturbation Conclusion 13 Understanding the protocol Understanding Read/Write protocol – Read and Write implementations in big distributed storage systems are not simple – Deconstruct the protocol structure • • • Which pieces are involved? Where data is sent to? Data reliably stored, mirrored, striped? 14 Observing Write Protocol Deconstruct protocol using passive observation – Run a series of write workload – Observe network and disk traffic – Correlation tools: convert traces into protocol structure EMC Centera Client write( ) Access Nodes Storage Nodes 15 Observation Results Client Access Node Write Req. R Primary SN Secondary SN – Phase 1: Write request establishment – Phase 2: Data transfer – Phase 3: Disk write, notify other SNs, commit – Phase 4: Series of acknowledgement TCP Setup Write Req R TCP Setup R Request Ack. Request Ack. Request Ack. Data Transfer Transfer Ack. SNx SNv SNw Write-Commit Write-Commit Write Complete Software ACKs time Software ACKs Object Write Protocol findings SNy Determine general properties – Primary SN handles generation of 2nd copy – Two new TCP connections / object write 16 Resolving Dependencies Cannot conclude dependencies from observation only – B after A != B depends on A • AN Must delay A, and see if B is delayed Primary SN Primary commit (pc) Secondary SN From observation only: Primary commit depends on secondary commit and sync. disk write Secondary Commit (sc) Conclude causality by delaying: -disk write traffic and -secondary commit 17 Delaying a Particular Message Need to delay a particular message – Leverage packet sizes – Modify NistNet • Delay specific message, not link – Ex: delay sc (90 bytes) Access Node Client Primary SN Secondary SN 299 bytes 509 161 289 161 509 161 375 321 Primary SN 321 CentraStar sc Linux TCP/UDP 539 Mod. NistNet 90 bytes prim. commit no 4 if size= 90 4 yes incoming packet 4 4 delay queue 18 Delaying secondary-commit AN Primary SN Secondary SN Secondary commit Resolving first dependency – Delay secondary commit primary commit also gets delayed Primary commit + delay Primary commit depends on the receipt of secondary commit 19 Delaying disk I/O traffic Delay disk writes at primary storage node Primary SN Primary SN Secondary-commit Primary-commit + Delay Disk Write CentraStar Pseudo-Dev ReiserFS disk req if WRITE no yes delay queue IDE Driver From observation and delay: Primary commit depends on secondary commit message and sync. disk write 20 Ability to analyze internal designs Intra-box techniques: Observation and perturbation by delay – Able to deduce Object Write protocol – Give ability to analyze internal design decisions Serial vs. Parallel Client AN – Primary SN handles the generation of 2nd copy (Serial) vs. Client AN handles both 1st and 2nd (Parallel) – EMC Centera: write throughput is more important – Decrease load on access nodes – increase write throughput 1 SN1 2 SN2 1 SN1 2 SN2 AN New TCP connections (internally) / object write – vs. using persistent connection to remove TCP setup cost – Prefer simplicity – no need to manage persistent conn. for all requests 21 Outline Introduction EMC Centera Overview Deducing Protocol Inferring Policies – Various system perturbation Conclusion 22 Inferring internal policies Write policies – Level of replication, Load balancing, Caching/buffering Read policies – Caching, Prefetching, Load balancing Try to infer – Is particular policy implemented? – At which level it is being implemented? • Ex: Read Caching at Client, Access Node, Storage Node? 23 System Pertubation Client Perturb the system – Delay and extra load write() Access Node 4 common loadbalancing factors: – CPU load • High priority while loop – Disk load • Background file copy + net delay ? SN 1 ? SN 2 ? ? SN 3 SN … Active TCP CPU CPU CPU CPU – Active TCP connection – Network delay 24 Write Load Balancing What factors determined which storage nodes are selected? Experiment: – Observe which primary storage nodes selected – Without load: writes are balanced – With load: writes skew toward unloaded nodes ? sn#1 Unloaded ? sn#2 sn#2Unloaded Loaded AN 25 Write Load Balancing Results Normal No Perturb Additional CPU Load Disk Load Network Load Incoming Net. Delay sn#1 sn#1 sn#1 sn#1 sn#1 sn#2 +CPU +Disk +TCP +Delay 26 Summary of findings Write Policies Replication Load balancing Write buffering Caching Two copies in two nodes attached to different power (reliability) EMC Centera: CPU usage (locally observable status) Network status is not incorporated Simplicity and Read Policies Storage node only (commodity filesystem) Reliability Access node and client does not cache. Storage nodes write synchronously Prefetching Storage node only (commodity filesystem) Access node and client does not prefetch Load Balancing Not implemented in earlier version Still reads from busy nodes 27 Conclusion Intra-box: – Observe and perturb – Deconstruct protocol and infer policies – No access to source code Power of probe points – More observation places – Ability to control the system Systems built with more externally visible probe points – Systems more readily understood, analyzed, and debugged – Higher-performing, more robust and reliable computer systems 28 Questions? 29