Locking Protocol Simulator Checkpoint Meeting

1
Locking Protocol Simulator Checkpoint Meeting

• Hongfei Guo Pradeep Tamma

2
Meeting Agenda

• Cache-master implementation issues
• Transaction processing assumptions
• Dealing with currency checking failure
• Modeling network cost
• Generating copier transactions
• Locking protocol
• Dealing with copier transactions
• Cache-master case experiment design
• Experiment results for single-site case
• See attached PDF document

3
Transaction Processing (master)

• Before a read-write Xact commits, push all writes
  and the Xact timestamp into the copier_queue
• (Need to be clarified)--Yes
• Commit time currency checking if fails, abort
  the Xact and restart it

4
Transaction Processing (Cache)

• For read-only user Xacts
• All the reads are done at the cache
• Commit time checking if there are reads that
  fail currency checking, abort the Xact and send
  it to the Master
• For read-write user Xacts
• All the reads are done at the cache
• After reads, shift the Xact to the master
  (together with info about the reads)
• Commit-time checking is performed at the master

5
Modeling the Network Cost

• Two more model parameters
• network_delay_message the cost of sending a
  remote query
• network_delay_object the cost of sending an
  object over the network
• Total network cost for a remote query asking for
  n objects
• Total network cost network_delay_message n
  network_delay_object
• We assume infinite bandwidth no waiting for
  network request

6
Generating Copier Transactions

• (Need to be clarified)--Yes
• The copier process wakes up periodically, empties
  the copier_queue, and submits a copier Xact to
  the cache
• Note that we wrap writes from a group of Xacts
  into a single copier Xact.
• Correction we send one transaction at a time,
  this way, we reduce the false negative due to
  currency checking failure
• We dont model log sniffing cost

7
Locking Protocol

• For user transactionsRead-locks short duration
• Write-locks long duration
• For copier transactions
• Read-locks none (Needs to be clarified)--Yes
• Write-locks short duration--Yes

8
Dealing with Copier Transactions

• Do we want to give higher priority to a copier
  Xact? If so, how aggressive?
• In order to improve local data quality?
• An example A copier xtact requests all locks
  simultaneously (detail?)
• NO!

9
Experiments

• Reference protocol
• Cache BSA
• Master BSA
• Additional performance metrics
• Parameters
• High level goals
• More detailed experiment settings

10
Additional Performance Metric

• Workload shift ratio the ratio of successful
  reads at the cache

11
Parameters

• Model parameters (fixed)
• Caching-related parameters
• Number of caches
• Currency bounds
• Refresh intervals
• Write probability
• Workload originated from the master

12
High Level Goals

• We want to study
• The influence of the number of caches
• The influence of write probability
• The influence of refresh interval
• The influence of currency bound

13
Influence of the Number of Caches

• Parameter setting
• Everything else being equal, vary the number of
  caches (use continuous refresh)
• Different curves correspond to different write
  probability
• Expectations
• When write probability is low, increasing the
  number of caches should increase the throughput
• When write probability is high, refresh cost
  might dominate
• In the rest of the experiments, we use only one
  cache.

14
Influence of Write Probability

• Parameter setting
• Everything else being equal, vary the write
  probability
• Different curves correspond to different refresh
  intervals and currency bounds
• Expectations
• Throughput decreases when write probability
  increases.
• With larger currency bounds, longer refresh
  intervals, the throughput deterioration should be
  slower.

15
Influence of Refresh Interval

• Parameter setting
• Everything else being equal, vary the refresh
  interval (0, B (currency bound), 2B, , infinity
  )
• Different curves correspond to different write
  probabilities
• Expectations
• Workload shift ratio drops when refresh interval
  increases beyond B
• Local workload ratio does not change much when
  refresh interval varies between 0-B
• Higher write probability causes lower throughput

16
Influence of Currency Bounds

• Parameter setting
• Everything else being equal, vary the currency
  bound
• Different curves correspond to different write
  probabilities and refresh intervals
• Expectation
• Larger currency bound causes higher workload
  shift ratio
• Shorter refresh interval causes higher workload
  shift ratio
• Lower write probability causes higher throughput

17
Issues

• Model parameter setting
• Symmetric setting for cache and master?
• Resource units is cache as powerful as master?
• Percentage of master originated workload?
• Cache-related parameter
• Should we separate read-only scenario from
  read-write scenario at the cache?