Mobility

Presented by:Mohamed Elhawary

Mobility Distributed file systems increase

availability Remote failures may cause serious

troubles Server replication (higher quality) Client replication, inferior, needs periodic

revalidation Cache coherence to combine

performance, availability and quality

Mobility (Cont’d)

s tron glocks

low ava ilab ility

con n ec ted w eak ly con n ec tedap p lica tion ad ap ta tion

d a ta exch an g e op era tion exch an g e

op tim is tic p ess im is tic

d iscon n ec ted

w eak lycach e coh eren ceh ig h ava ilab ility

con s is ten tClient server as CodaPairwise as BayouTwo level hierarchy as Oracle

Disconnected Operation in Coda Enable clients to continue accessing

data during failures Cashing data on the clients with

right back upon reconnection Designed for a large num. of clients

and a smaller num. of servers Server replication and call back

break Preserve transparency

Scalability Whole file cashing, cash miss on

open only Simple but involve communication

overhead Placing functionality on the clients

unless violating security or integrity Avoids system wide rapid change

Optimistic replica control Disconnection is involuntary Extended disconnections Leases place time bound on locks

but disconnected client loses control after expiration

Low degree of write sharing in Unix Applied to server replication for

transparency

Client structure Venus is a user

level process Works as a cache

manager

Venus States Hoarding state:

normal operation relying on server replication

Reintegration state: resynchronize its state

Venus can be in different states with respect to different volumes

Hoarding Hoard useful data in anticipation of

disconnection Combines implicit and explicit source of

information File reference behavior can’t be predicted Disconnections are unpredictable Priority of cached objects is a function of

user given priority (unsafe) and recent usage (hard to manage)

Hierarchical cache management

Hoard walking Periodically restores cache

equilibrium Unexpected disconnection before a

scheduled walk is a problem Revaluate name bindings Revaluate priorities of all entries Solve the callback break problem Policies for directories and files

Emulation Venus functions as a pseudo server Generates temporary file identifiers Behaves as faithful as possible, but no

guarantee Cache entries of modified assume infinite

priority, except when cache is full Use a replay log Use nonvolatile storage Meta data changes through atomic

transactions

Reintegration Venus propagates changes made in

emulation Update suspended in the volume being

integrated until completion, it may take time!

Replace temporary ids with permanent ones, but may be avoided

Ship replay log Execute replay as a single transaction,

is that fair?

Replay algorithm Log is parsed Log is validated using store id Store create a shadow file and data

transfer is deferred Data transfer Commits and release locks In case of failures, user can replay

selectively Using dependence graphs may enhance

reintegration

Evaluation

Evaluation (cont’d)

Evaluation (cont’d)

Flexible update propagation in Bayou Weekly consistent replication can

accommodate propagation policies Don’t assume client server as in Coda Exchange update operations and not

the changes, suite low B/W networks Relies on the theory of epidemics Storage vs. bandwidth

Basic Anti-entropy Storage consists of an ordered log of

updates and a database Two replicas bring each other up-to-date by

agreeing on the set of writes in their logs Servers assigns monotonically increasing

accept stamp to new writes Write propagates with stamps and server id Partial accept order to maintain the prefix

property Needs to be implemented over TCP

Basic Anti-entropy (Cont’d)

Effective write-log Each replica discard stable writes as a

pruned prefix from write log They may have not fully propagated

and may require full database transfer Storage bandwidth tradeoff Assigns monotonically increasing CSN to

committed writes and infinity to tentative writes

Write A precedes write B if…

Anti-entropy with com. writes

Write log truncation To test if a server is missing writes,

S.O version vector Penalty involve database transfer

and roll back

Write log truncation (Cont’d)

Session guarantees Causal order as a refinement of

accept order Each server maintains a logical

clock that advances with new writes

Total order through <CSN, accept order, server id>

Light weight creation and retirement Versions vectors are updated to

include or exclude servers Creation and retirement are handled

as a write that propagates via anti-entropy

<Tk,i , Sk> is Si ’s server id Tk,i + 1 initialize accept stamp counter Linearly creation increase id’s and

version vectors dramatically

Disadvantages When num of replicas is larger

than update rate, cost of exchanging version vectors is high

If the update rate is larger than the commit rate, the write log will grow large

Policies When to reconcile Which replicas to reconcile How to truncate the write log Selecting a server to create a new

replica Depend on network characteristics

and threshold values

Evaluation Experiments were taken for an e-mail

application, how far is that really represent the general dist. File system

Only committed writes are propagated Sparc running SunOS (SS) 486 laptops running linux (486) 3000 and 100 byte messages between 3

replicas 520 public key overheads and 1316 bytes

update schema and padding

Execution vs. writes propagated

Execution for 100 writes

Network independent

Effect of server creations When servers created from initial

one, 20K version vector for 1000 servers

When linearly created, 4 MB version vectors for 1000 servers

Time to test if a write is among those represented by version vector may grow cubically if linearly created

Conclusion Disconnected operation through

Cedar and PCMAIL is less transparent than Coda

Golding’s time stamped anti entropy protocol is similar to bayou, with heavier weight mechanism to create replicas and less aggressive to claim storage