COMP 655: Distributed/Operating Systems Summer 2011 Dr. Chunbo Chu Week 7: Fault Tolerance 11/13/20151Distributed Systems - COMP 655
11/13/2015Distributed Systems - COMP 6552 Fault Tolerance Fault tolerance concepts Implementation – distributed agreement Distributed agreement meets transaction processing: 2- and 3-phase commit Bonus material Implementation – reliable point-to-point communication Implementation – process groups Implementation – reliable multicast Recovery Sparing
11/13/2015Distributed Systems - COMP 6553 Fault tolerance concepts Availability – can I use it now? –Usually quantified as a percentage Reliability – can I use it for a certain period of time? –Usually quantified as MTBF Safety – will anything really bad happen if it does fail? Maintainability – how hard is it to fix when it fails? –Usually quantified as MTTR
11/13/2015Distributed Systems - COMP 6554 Comparing nines 1 year = 8760 hr Availability levels –90% = 876 hr downtime/yr –99% = 87.6 hr downtime/yr –99.9% = 8.76 hr downtime/yr –99.99% = min downtime/yr –99.999% = min downtime/yr
11/13/2015Distributed Systems - COMP 6555 Exercise: how to get five nines 1.Brainstorm what you would have to deal with to build a single-machine system that could run for five years with 25 min downtime. Consider: –Hardware failures, especially disks –Power failures –Network outages –Software installation –What else? 2.Come up with some ideas about how to solve the problems you identify
11/13/2015Distributed Systems - COMP 6556 Multiple machines at 99% Assuming independent failures
11/13/2015Distributed Systems - COMP 6557 Multiple machines at 95% Assuming independent failures
11/13/2015Distributed Systems - COMP 6558 Multiple machines at 80% Assuming independent failures
11/13/2015Distributed Systems - COMP ,000 components
11/13/2015Distributed Systems - COMP Things to watch out for in availability requirements What constitutes an outage … –A client PC going down? –A client applet going into an infinite loop? –A server crashing? –A network outage? –Reports unavailable? –If a transaction times out? –If 100 transactions time out in a 10 min period? –etc
11/13/2015Distributed Systems - COMP More to watch out for What constitutes being back up after an outage? When does an outage start? When does it end? Are there outages that don’t count? –Natural disasters? –Outages due to operator errors? What about MTBF?
11/13/2015Distributed Systems - COMP Ways to get 99% availability 1.MTBF = 99 hr, MTTR = 1 hr 2.MTBF = 99 min, MTTR = 1 min 3.MTBF = 99 sec, MTTR = 1 sec
11/13/2015Distributed Systems - COMP More definitions failure error fault causes may cause Fault tolerance is continuing to work correctly in the presence of faults. Types of faults: transient intermittent permanent
11/13/2015Distributed Systems - COMP Types of failures
11/13/2015Distributed Systems - COMP If you remember one thing Components fail in distributed systems on a regular basis. Distributed systems have to be designed to deal with the failure of individual components so that the system as a whole –Is available and/or –Is reliable and/or –Is safe and/or –Is maintainable depending on the problem it is trying to solve and the resources available …
11/13/2015Distributed Systems - COMP Fault Tolerance Fault tolerance concepts Implementation – distributed agreement Distributed agreement meets transaction processing: 2- and 3-phase commit
11/13/2015Distributed Systems - COMP Two-army problem Red army has 5,000 troops Blue army and White army have 3,000 troops each Attack together and win Attack separately and lose in serial Communication is by messenger, who might be captured Blue and white generals have no way to know when a messenger is captured
11/13/2015Distributed Systems - COMP Activity: outsmart the generals Take your best shot at designing a protocol that can solve the two-army problem Spend ten minutes Did you think of anything promising?
11/13/2015Distributed Systems - COMP Conclusion: go home “agreement between even two processes is not possible in the face of unreliable communication”
11/13/2015Distributed Systems - COMP Byzantine generals Assume perfect communication Assume n generals, m of whom should not be trusted The problem is to reach agreement on troop strength among the non-faulty generals
11/13/2015Distributed Systems - COMP Byzantine generals - example n = 4, m = 1 (units are K-troops) (a)Multicast troop-strength messages (b)Construct troop-strength vectors (c)Compare notes: majority rules in each component Result: 1, 2, and 4 agree on (1,2,unknown,4)
11/13/2015Distributed Systems - COMP Doesn’t work with n =3, m =1
11/13/2015Distributed Systems - COMP Fault Tolerance Fault tolerance concepts Implementation – distributed agreement Distributed agreement meets transaction processing: 2- and 3-phase commit
11/13/2015Distributed Systems - COMP Distributed commit protocols What is the problem they are trying to solve? –Ensure that a group of processes all do something, or none of them do –Example: in a distributed transaction that involves updates to data on three different servers, ensure that all three commit or none of them do
11/13/2015Distributed Systems - COMP phase commit CoordinatorParticipant What to do when P, in READY state, contacts Q
11/13/2015Distributed Systems - COMP If coordinator crashes Participants could wait until the coordinator recovers Or, they could try to figure out what to do among themselves –Example, if P contacts Q, and Q is in the COMMIT state, P should COMMIT as well
11/13/2015Distributed Systems - COMP phase commit What to do when P, in READY state, contacts Q If all surviving participants are in READY state, 1.Wait for coordinator to recover 2.Elect a new coordinator (?)
11/13/2015Distributed Systems - COMP phase commit Problem addressed: –Non-blocking distributed commit in the presence of failures –Interesting theoretically, but rarely used in practice
11/13/2015Distributed Systems - COMP phase commit CoordinatorParticipant
11/13/2015Distributed Systems - COMP Bonus material Implementation – reliable point-to- point communication Implementation – process groups Implementation – reliable multicast Recovery Sparing
11/13/2015Distributed Systems - COMP RPC, RMI crash & omission failures Client can’t locate server Request lost Server crashes after receipt of request Response lost Client crashes after sending request
11/13/2015Distributed Systems - COMP Can’t locate server Raise an exception, or Send a signal, or Log an error and return an error code Note: hard to mask distribution in this case
11/13/2015Distributed Systems - COMP Request lost Timeout and retry Back off to “cannot locate server” if too many timeouts occur
11/13/2015Distributed Systems - COMP Server crashes after receipt of request Possible semantic commitments –Exactly once –At least once –At most once NormalWork doneWork not done
11/13/2015Distributed Systems - COMP Behavioral possibilities Server events –Process (P) –Send completion message (M) –Crash (C) Server order –P then M –M then P Client strategies –Retry every message –Retry no messages –Retry if unacknowledged –Retry if acknowledged
11/13/2015Distributed Systems - COMP Combining the options
11/13/2015Distributed Systems - COMP Lost replies Make server operations idempotent whenever possible Structure requests so that server can distinguish retries from the original
11/13/2015Distributed Systems - COMP Client crashes The server-side activity is called an orphan computation Orphans can tie up resources, hold locks, etc Four strategies (at least) –Extermination, based on client-side logs Client writes a log record before and after each call When client restarts after a crash, it checks the log and kills outstanding orphan computations Problems include: –Lots of disk activity –Grand-orphans
11/13/2015Distributed Systems - COMP Client crashes, continued More approaches for handling orphans –Re-incarnation, based on client-defined epochs When client restarts after a crash, it broadcasts a start-of-epoch message On receipt of a start-of-epoch message, each server kills any computation for that client –“Gentle” re-incarnation Similar, but server tries to verify that a computation is really an orphan before killing it
11/13/2015Distributed Systems - COMP Yet more client-crash strategies One more strategy –Expiration Each computation has a lease on life If not complete when the lease expires, a computation must obtain another lease from its owner Clients wait one lease period before restarting after a crash (so any orphans will be gone) Problem: what’s a reasonable lease period?
11/13/2015Distributed Systems - COMP Common problems with client-crash strategies Crashes that involve network partition (communication between partitions will not work at all) Killed orphans may leave persistent traces behind, for example –Locks –Requests in message queues
11/13/2015Distributed Systems - COMP Bonus material Implementation – reliable point-to- point communication Implementation – process groups Implementation – reliable multicast Recovery Sparing
11/13/2015Distributed Systems - COMP How to do it? Redundancy applied –In the appropriate places –In the appropriate ways Types of redundancy –Data (e.g. error correcting codes, replicated data) –Time (e.g. retry) –Physical (e.g. replicated hardware, backup systems)
11/13/2015Distributed Systems - COMP Triple Modular Redundancy
11/13/2015Distributed Systems - COMP Tandem Computers TMR on –CPUs –Memory Duplicated –Buses –Disks –Power supplies A big hit in operations systems for a while
11/13/2015Distributed Systems - COMP Replicated processing Based on process groups A process group consists of one or more identical processes Key events –Message sent to one member of a group –Process joins group –Process leaves group –Process crashes Key requirements –Messages must be received by all members –All members must agree on group membership
11/13/2015Distributed Systems - COMP Flat or non-flat?
11/13/2015Distributed Systems - COMP Effective process groups require Distributed agreement –On group membership –On coordinator elections –On whether or not to commit a transaction Effective communication –Reliable enough –Scalable enough –Often, multicast –Typically looking for atomic multicast
11/13/2015Distributed Systems - COMP Process groups also require Ability to tolerate crash failures and omission failures –Need k+1 processes to deal with up to k silent failures Ability to tolerate performance, response, and arbitrary failures –Need 3k+1 processes to reach agreement with up to k Byzantine failures –Need 2k+1 processes to ensure that a majority of the system produces the correct results with up to k Byzantine failures
11/13/2015Distributed Systems - COMP Bonus material Implementation – reliable point-to- point communication Implementation – process groups Implementation – reliable multicast Recovery Sparing
11/13/2015Distributed Systems - COMP Reliable multicasting
11/13/2015Distributed Systems - COMP Scalability problem Too many acknowledgements –One from each receiver –Can be a huge number in some systems –Also known as “feedback implosion”
11/13/2015Distributed Systems - COMP Basic feedback suppression in scalable reliable multicast If a receiver decides it has missed a message, it waits a random time, then multicasts a retransmission request while waiting, if it sees a sufficient request from another receiver, it does not send its own request server multicasts all retransmissions
11/13/2015Distributed Systems - COMP Hierarchical feedback suppression for scalable reliable multicast messages flow from root toward leaves acks and retransmit requests flow toward root from coordinators each group can use any reliable small- group multicast scheme
11/13/2015Distributed Systems - COMP Atomic multicast Often, in a distributed system, reliable multicast is a step toward atomic multicast Atomic multicast is atomicity applied to communications: –Either all members of a process group receive a message, OR –No members receive it Often requires some form of order agreement as well
11/13/2015Distributed Systems - COMP How atomic multicast helps 1.Assume we have atomic multicast, among a group of processes, each of which owns a replica of a database 2.One replica goes down 3.Database activity continues 4.The process comes back up 5.Atomic multicast allows us to figure out exactly which transactions have to be re-played (see pp )
11/13/2015Distributed Systems - COMP More concepts Group view View change Virtually synchronous –Each message is received by all non-faulty processes, or –If sender crashes during multicast, message could be ignored by all processes
11/13/2015Distributed Systems - COMP Virtual synchrony picture Basic idea: in virtual synchrony, a multicast cannot cross a view-change
11/13/2015Distributed Systems - COMP Receipt vs Delivery Remember totally-ordered multicast …
11/13/2015Distributed Systems - COMP What about multicast message order? Two aspects: –Relationship between sending order and delivery order –Agreement on delivery order Send/delivery ordering relationships –Unordered –FIFO-ordered –Causally-ordered If receivers agree on delivery order, it’s called totally-ordered multicast
11/13/2015Distributed Systems - COMP Unordered Process P1Process P2Process P3 sends m1 sends m2 delivers m1 delivers m2 delivers m1
11/13/2015Distributed Systems - COMP FIFO-ordered Agreement on: m1 before m2 m3 before m4 Process P1Process P2Process P3 sends m1 sends m2 delivers m1 delivers m3 delivers m2 delivers m4 delivers m3 delivers m1 delivers m2 delivers m4 Process P4 sends m3 sends m4
11/13/2015Distributed Systems - COMP Six types of virtually synchronous reliable multicast Relationship between sending order and delivery order Agreement on delivery order
11/13/2015Distributed Systems - COMP Implementing virtual synchrony Don’t deliver a message until it’s been received everywhere - but “everywhere” can change (a)7’s crash is detected by 4, which sends a view-change message (b)Processes forward unstable messages, followed by flush (c)When have flush from all processes in new view, install new view
11/13/2015Distributed Systems - COMP Bonus material Implementation – reliable point-to- point communication Implementation – process groups Implementation – reliable multicast Recovery Sparing
11/13/2015Distributed Systems - COMP Recovery from error Two main types: –Backward recovery to a checkpoint (assumed to be error-free) –Forward recovery (infer a correct state from available data)
11/13/2015Distributed Systems - COMP More about checkpoints They are expensive Usually combined with a message log Message logs are cleared at checkpoints Recovering a crashed process: –Restart it –Restore its state to the most recent checkpoint –Replay the message log
11/13/2015Distributed Systems - COMP Recovery line == most recent distributed snapshot
11/13/2015Distributed Systems - COMP Domino effect
11/13/2015Distributed Systems - COMP Bonus material Implementation – reliable point-to- point communication Implementation – process groups Implementation – reliable multicast Recovery Sparing
11/13/2015Distributed Systems - COMP Sparing Not really fault tolerance But it can be cheaper, and provide fast restoration time after a failure Types of spares –Cold –Hot –Warm The spare may or may not also have regular responsibilities in the system
11/13/2015Distributed Systems - COMP Switchover Repair is accomplished by switching processing away from a failed server to a spare
11/13/2015Distributed Systems - COMP Questions on switchover Has the failed system really failed? Is the spare operational? Can the spare handle the load? –May need a way to block medium to low priority work during switchovers How will the spare get access to the failed server’s data? What client session data will be preserved, and how?
11/13/2015Distributed Systems - COMP More switchover questions What about configuration files? What about network addressing? What about switching back after the failed server has been repaired? –Partial shutdown of the spare –Updating directories to redirect part of the load –Making up for lost medium-to-low priority work