Skip to content

Latest commit

 

History

History
389 lines (356 loc) · 17.2 KB

Projects.page

File metadata and controls

389 lines (356 loc) · 17.2 KB

% Research Projects % Yang Zhang

  • key
    • QUESTION: means a question that Sam may be able to help answer
    • TODO: means not done yet (not necessarily that I'll follow up on these, though)

H-Store Recovery and Elasticity

H-Store

  • distributed transaction manager is exciting and central to H-Store
  • H-Store as a whole easy to believe in and is not a marginal project
  • recovery/elasticity could make for interesting demos

Recovery

  • recovery: based on HARBOR
    • HARBOR's approach:
      • asynchronous disk checkpointing
      • query snapshots for updates since failure
      • query current data for final updates since start of recovery
      • generally: snapshot isolation/time travel minimizes impact on availability
    • obvious strategies
      • network-only: recover entirely from replicas (network BW > disk BW, but interference with ongoing transaction processing)
      • hybrid network and disk: similar to HARBOR, but taking snapshots is hard/expensive
      • disk-only: same as ARIES
  • thought process
    • assumptions
      • high update rate
      • updates distributed uniformly at random
      • so all pages are quickly touched
      • assuming that all updates are applied in same order on each replica; is this valid?
    • stupid algo
      • regularly flush page snapshots to disk (asyncly)
        • use COW to take snapshots/avoid blocking
        • COW lasts only for duration of flush; no copy might be nec
        • if update during flush is likely, then just make eager copy
      • recover by seeing which pages are touched and grab latest from replicas
      • but all pages are out of date (useless)
      • so you'd need to transfer entire state from other machines, a long op
        • to avoid interrupting other replicas, transfer snapshots
        • simultaneously stream in new updates and apply them after getting state
    • consider logging
      • concerned only with redo logging (to prevent lost updates)
      • volatile undo logging: we need this anyway to be able to undo xacts
        • so one log for both purposes
        • redo log should generally dictate length
        • xacts are short, so undos don't stick around
      • non-vol undo logging: for buffer frame steals that swap dirty pages to disk
      • but we have no disk, and no such thing as dirty pages on replicas
    • algo v1: assuming identical update ordering (see above assumption)
      • LSN: log seq num
      • an update is a lower-level concept than query updates, transactions, commits
      • we log all updates; hence we log undo operations as well
      • each page has pageLSN of latest touching update
      • minLSN is the min pageLSN
      • in normal operation, do regular flushing as above
      • on recovery, ask others for log of all updates following minLSN
        • let these be the redo updates
      • simultaneously stream in new updates that have been submitted to this partition
        • note that these updates are higher-level updates
        • let these be the new updates
        • let minNewLSN be the LSN of the first new update
        • ensure that you've obtained redo updates up through minNewLSN - 1
        • apply new updates after redo updates
        • skip updates to pages older than its pageLSN (or just re-apply them if log records log idempotent ops, e.g. "after" images)
      • never quiesces the system
    • truncating logs
      • when any one replica fails at minLSN, others better have logs starting at or before minLSN
        • if replica is dead for long time, start flushing logs to disk
        • alternatively, give up on logs that grow too large, and have the recovering node discard and fetch all state (no logs involved)
      • also may want to flush to disk anyway to minimize memory storage requirements
      • normal operation: keep others informed of your minLSN regularly (e.g. as you flush each page)
        • each replica can then truncate up to min of all received minLSNs
    • vs. HARBOR, which maintains no logs
      • however, historical queries effectively "generate the log"
      • minLSN = HWM
      • time-travel + historical queries seem to be mainly useful for working with non-replicas (containing different logical partitions of objects; i.e. multiple recovery objects and recovery buddies per object)
      • also, this may be faster if many tuples is inserted then deleted during the downtime (not sure if HARBOR does this, but can modify query to ignore such tuples)
      • different assumptions
        • HARBOR has abundant storage capacity (disk), so can afford to keep history
        • overhead & complexity in snapshotting and maintaining segments
        • QUESTION: did HARBOR also consider using logging? what were the main reasons for current design?
      • QUESTION: why perform final updates? why not wait for the highest epoch to expire, and stream in queries from coord, then apply them after recovery, rather than introducing quiescing?
      • must synchronously advance timestamps (with coord)
  • another version of algo
    • round-robin full-database snapshotting among the replicas
      • enables xact logging
  • decision tree
    • for multi-partition xacts: seq good enough?
      • yes: h-store default
      • no: pre-order or post-order?
        • pre-order: speculation (of success vs. failure, not ordering)
        • post-order: pessimistic or optimistic concurrency control?
          • pessimistic: distributed locking
          • optimistic: distributed read sets and write sets
    • unclear which of pre-order, pessimistic, optimistic is better
  • QUESTION: how to evaluate performance?
    • it's possible the recovering node may never catch up if transactions are being processed at maximum (resource-saturating) throughput
    • possibility: how much slower is the system's peak throughput
  • TODO look at
    • Paxos made live: snapshots in section 5.5, which may be useful to think about the interface between the replication system and the database
  • other thoughts
    • Sam thinks this is the most concrete idea so far

Elasticity

  • elasticity: seamlessly add/remove nodes into the system while minimizing impact on availability
  • may be closely related to recovery
  • how much churn?
    • Google-style incremental growth and frequent recovery
    • pay-as-you-go services encourage dynamically resizing (perhaps substantially) throughout the day
  • other thoughts
    • elasticity slightly more exciting than recovery

H-Store distributed transaction manager: transaction consistency among replicas XXX ADDRESS TODOS

  • high-level picture

    • distributed xact mgr
      • Coordinator (perhaps a library on each client, like JDBC) interacts with DB Master Executor
        • Coord sends xact blobs to ME
        • ME inspects xact, identifies partitions to Coord
        • Coord dispatches queries to partitions
        • within each partition, master waits to hear from majority of nodes (incl. self)
        • Coord sends results to the ME; ME processes them
        • single-partition xacts commit immediately while in the partition
          • Coord failures don't matter; partition exclusively holds the state of the xact
        • multi-partition xacts require Commit Log servers
          • ME notifies Coord of commit/abort
          • Coord records results to the cluster of Commit Log servers
          • these are to deal with Coord failure (only relevant in the case of multi-partition xacts, since single-partition xacts commit immediately)
          • actual xact state is recorded here among the servers (also using consensus/master-slave hybrid)
      • partitions use hybrid strategy
        • use paxos consensus to decide on master lease
        • master-slave replication
      • recovery and elasticity deeply intertwined with distributed xact mgr
      • open question: wide-area replication vs. relaxed consistency in the event of site failures
        • how much does wide-area replication actually hurt throughput?
      • open question: symmetric vs. asymmetric sites?
      • to conserve resources/reduce cost of availability
        • ship logs to dest nodes; they only store, no computations required
        • what about chain-replication?
    • one problem with elasticity: re-balancing
  • primary/backup replication aka master/slave replication

    • operation
      • primary server receives all requests and broadcasts them to backup replicas
      • after master receives acks from majority of replicas, ack client
    • properties
      • master orders all requests, solving the ordering problem
      • handling master failures is harder
      • if network partitions or behaves arbitrarily, need to either use consensus or totally ordered broadcast
  • consensus (Paxos)

    • tolerates only $\lfloor {\frac{n-1}{2}} \rfloor$ failures
    • so, on partition, only one partition can continue operation
    • strong consistency and partition tolerance, but sacrifices availability
  • hybrid primary/backup with consensus

    • use paxos to decide who gets the master lease
    • advantage: within a primary/backup group, more failures can be tolerated
    • disadvantage: master must stay alive, or need to re-lease
    • used by GFS and BigTable
  • totally ordered broadcast

    • deliver msgs to a group of processes in the same order
    • shown to be equiv. to consensus, if strong consistency is required
    • with weaker semantics, other solutions possible
      • there exists an algo that allows group to continue functioning even in case of partition
      • leads to fork consistency, where each partition works on its own copy of the data
      • merging forked copies is hard, slow
        • how to handle conflicting transactions?
  • distributing transactions

    • single partition, read-only: any replica; attach last xact ID
    • single partition, read-write: all replicas
    • one-shot, no aborts: TODO cannot parse: "order components of xact relative to overlapping multiple partition xacts"
    • general transactions
      • if aborts are possible, use traditional concurrency control; slows down all xacts in that partition
      • 2PC required to agree on committed/aborted state of each xact
      • batching may reduce time spent in this mode
  • centralized global coord: TODO don't grok

  • viewstamped replication: TODO

Alternative Data Stores

Eddie Kohler and Robert Morris want to create a data store that is simpler (and hopefully more usable) than a conventional DBMS.

  • hot area: Dynamo, HP SOSP 2008, etc.
  • "ACID unnecessary"...but what semantics do you want?
  • probably a key-value store with get()/put() interface, a la BerkeleyDB
  • one concrete idea: a simple log storage system
    • every so often, collect/compress/sort/optimize into a nicer data structure
    • log is the only place where recent updates are stored
    • also don't worry about durability
  • other thoughts
    • very amorphous
    • not terribly excited about disk
  • TODO: related work, think about a more concrete premise

Object Database

  • something in the context of H-Store?
  • QUESTION: what exactly are the performance issues with ORMs?
  • other thoughts
    • I like hacking on my stupid personal object database system

How to make OLTP and OLAP co-exist?

  • original problem: Pawan had problems running OLAP queries on web app DB
  • solution: use a DBMS with weaker consistency (e.g. snapshot isolation). duh
    • HARBOR applies this principle, in a sense

Multicore Database Projects

Build a multicore OLAP system (i.e. relational Phoenix)

  • Phoenix ignores I/O
  • you can try to attach enough disks until you saturate the CPU, but...
    • today it's all about horizontal scaling with commodity hardware
      • cabinets with disk arrays are dead because they're disproportionately more expensive, harder to replace, etc.
      • also, high-profile success stories (Google, etc.) advocate scaling out with commodity
    • one way to spin this project: "we reduce your entire data warehouse onto a single chip (well, system)!"
      • one question is whether people will buy this, or stick to their cheap commodity guns
      • QUESTION: exactly how much more TCO are disk array systems, really?
    • what are the stats? what's the I/O bandwidth limit? how many disks?
      • maxhd: "the fastest modern desktop hard drives transfer data at a maximum of about 120 MB/s"
      • dbcol: "Despite capacities and transfer rates increasing by factors or 10,000 and 100 respectively, typical drive ETB has actually decreased by a factor of 100."
      • tilera: "Up to 50 Gbps of I/O bandwidth"
        • 50,000Mbps / 120Mbps = 416 disks, so more than enough I/O bandwidth
        • this is a minimum; with anything less than perfectly sequential access, must attach even more
    • QUESTION: is OLAP ultimately bound by disk capacity or disk bandwidth?
      • i.e., do people add more (smaller) disks/systems to decrease their query times? or is the number of disks/systems entirely dictated by capacity?
    • TODO: what have people done already in multiprocessor/multicore OLAP?
      • certainly, big DBMS vendors must already have multiprocessor operators in their products, no? then need to have a different story
  • what about using distributed main memory as storage?
    • esp. with the arrival of 10gigE, much faster than disks
      • 50 GB/s IO bandwidth, so 50GB/(10Gb/8)=40 10gigE network cards
      • TILE64 already have dual XAUI
    • but: OLAP generally deals with large data warehouses, so making memory the capacity bottleneck is not cost-effective (I assume warehouses are large)
  • other issues
    • research nugget? many of the tricks we've been discussing tend to be highly hardware specific
    • saturating the I/O bandwidth still doesn't mean anything; CPUs can still sit idle, which means no scalability
    • who knows what the architecture of tomorrow really will be? the stats quoted above (e.g., bandwidth) can vary substantially from architecture to architecture
    • people seem less excited about OLAP in general (parallel processing is somehow slightly sexier, esp. if it requires high availability, a la Map-Reduce)
    • this should really just be a C-Store sub-project, which would make this a marginal project

Speculatively execute transactions in multiple orders

  • premise so far
    • can't always partition transactions cleanly among tiles; this is for when working with shared memory
  • this is aimed at improving the performance of workloads that involve multiple concurrent transactions where there's blocking due to pess. concurrency control
    • so this won't do for strictly OLTP workloads
      • queries are too short and selective
      • so no H-Store
    • context must be either mixed OLTP/OLAP or strict OLAP
    • strict OLAP: multiple large datasets probably makes things harder
  • TODO: need to figure out a more concrete premise
    • it sounds from the premise above that we were thinking of H-Store, but its OLTP queries are short
    • so a conventional RDBMS? C-Store?
  • reservations
    • where would this be usable? need to cluster data enough such that they are not in the same tile but in the same host...
    • speculation comes at a cost (undo or redo, pick one); doubtful whether gains offset costs
    • (related to above point) wasteful: people will want to do other things on their multicore systems
    • marginal project

Speculative prediction

  • no idea here
  • no idea about domains/applications either
  • maybe a related question is what the analogous general programming model look like

Safe File-Sharing

Start with social networks as a trust network to avoid getting caught by the *AA. XXX FINISH

  • general guiding principles
    • want to find a lot of files
    • onion routing is slow; expected(min(xs)) is a decreasing function of |xs|
    • swarming is fast; this is another reason to want to maximize the number of files present
  • users publish pubkey to profile to restrict the scope of file searches
  • is searching hard? indexing seems hard...
    • supernode selection in an expander graph such that every span-$r$ subgraph contains at least one supernode
  • meet halfway by also doing partial dissemination, particularly of rare files
  • plausible denialibility: clique swarms
  • TODO: see also intel research's work on random walks in social networks
  • TODO: related/background work search!

Distributed Programming Models

  • harder than concurrent programming models, which itself is far from solved
  • focus on expression problem and/or program synthesis?
  • related work
    • process calculi
    • orchestration languages: orc
      • address issues like cancellation
    • awesome and underrated: P2
    • the pragmatic: mace, erlang, etc.
    • stream processing: Dryad, MapReduce, Borealis, maybe WaveScript
    • unknown: MPD, Acute
    • parallel PL:
      • X10
      • Chapel
      • Fortress
      • Titanium
      • Ease, Carnap
    • marginally: Mozart (aspects, mobile agents)
    • http://portal.acm.org/citation.cfm?id=800174.809772