1. Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement
Juan Rubio, Lizy K. John Charles Lefurgy
Laboratory for Computer Architecture IBM Austin Research Lab
The University of Texas at Austin, USA
Presented by Sean Leather
Laboratory for Computer Architecture
Good morning. Thank you for coming to this presentation. My name is _____ and I am graduate student at the University of Texas at Austin. Today I’ll be talking about …

2. Commercial Systems
Computer systems running commercial workloads operate on large amounts of data
Researchers have noticed that performance is hindered by data accesses
System architecture trends point to a distributed storage model with a non-uniform access latency for disk and memory
[go through this slide fast!]
Latency to access remote data can be an order of magnitude more than local data
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

3. Data Placement Goal: Challenges: Place data to reduce access penalties
It is difficult when looking at large amounts of data, a handful of processors and multiple operations
Uses of the data change with time
Key points:
Looking at ALL the possible combinations is hard
Even IF we figure out the right way, the usage patterns change
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

4. Data Placement: Example5 4 6 3 5 3 6 4
[present example]
For each figure, the dotted line represents a node. These nodes contain processors – represented by the squares on the left, and blocks of data, represented by the colorful blocks on the right.
The lines show the usage of those blocks of data by the processor.
The number on top of each line represents the amount of data that needs to be transferred.
Of these transfers, those that go beyond the boundaries of a node result in the longest latencies and therefore affect the performance the most.
We can see that by changing the location of some blocks (purple and green), the cost of those inter-node transfers can be reduced from 16 to 13.
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

5. Approach Static data placement Run-time data reorganization
Applied before the workload runs
Organizes blocks of data across disks of the system to result in low number of remote accesses
Run-time data reorganization
Applied while the system runs the workload
Used to adapt the layout of blocks of data to characteristics of the workload
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

6. Outline Problem Static data placement Run-time data reorganization
Simulated Annealing (SA)
Evaluation
Summary
This is an outline of the rest of the talk.
I just present an introduction and motivation behind our study.
Now, I’ll explain the simulated annealing process and the reason why we are using it in this work.
In the following sections I will explain how the simulated annealing technique can be used to guide the placement of data in a new system, and how to reorganize the data during run-time.
Then I will present an evaluation of these ideas, and will concludes the talk with a summary of our observations.
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

7. Static Data Placement
Arranges blocks of data across disks of the system
Approach:
Use probabilistic knowledge about the workload to formulate a cost function
Obtain layout that minimizes the cost function
Update disks to reflect the resulting layout
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

8. Run-time Data Reorganization
Approach:
Periodically run a reorganization routine for the whole system
Compute a cost function based on the description of some completed and pending operations
Determine changes to layout that would lower the cost function
Pre-fetch the data from the remote disks
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

9. Simulated Annealing [1/2]
Choose a number of “steps” (iterations)
For each step
Randomly introduce a perturbation (a small change to the current combination)
Always accept the new alternative if it reduces the cost
Randomly accept some alternatives that increase the cost (uphill change)
Slowly decrease the uphill acceptance probability
Later steps are less likely to accept bad perturbations
Simulated annealing is a heuristic commonly used to minimize the value of a cost function while avoiding a local minimum
An iterative algorithm
It uses a probabilistic approach to quickly achieve an adequate solution
This heuristic is used to minimize or reduce the value of the cost function as required by the previous 2 ideas.
To perform the algorithm, we must first select a number of steps to perform. We have to keep in mind that since this is an iterative process, the number of steps has a strong impact over the quality of solution.
A perturbation is a small change to the current system. And therefore it is specific to the application in question. For example in the examples I presented earlier, a perturbation might consist of moving a block of data from one node to another node.
Each step has an associated “temperature”, which is an abstract variable that affects the value of the uphill probability for a particular combination.
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

10. Simulated Annealing [2/2]
A randomizing algorithm, allows SA to quickly explore a vast design space
Accepting uphill changes allows SA to escape a local minima
Uphill changes can also be bad
We reduce the probability of accepting them as the exploration progresses
This uphill acceptance probability was modeled based on the physical annealing process. And it’s related to the temperature by this exponential expression.
Since the temperature is decreased between the different steps of the algorithm, the uphill acceptance probability also decreases. As a result, the uphill changes turn to be more conservative.
We can see that in this plot that describes the progress made by the simulated annealing algorithm to obtain the minimum cost for a sample function. The blue line represents the cost obtained by SA after each step, and the red curve represents the cost obtained by the Iterative Improvement method, which only accepts a perturbation when it reduces the cost of the objective function.
The advantage of SA come from the fact that it accepts some combinations that appear to hurt the solution. But turn to be beneficial at the long term.
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

11. System Full-system simulation (SimOS-PPC) 4 x 4 cc-NUMA
1 GHz CPUs, 512 MB per node, 7 disk units per node, 128-bit 100 MHz bus
128 bit 200 MHz inter-node bus
Directory-based cc-NUMA
System runs AIX 4.3.1
Using simulation in this work allowed me to measure characteristics of the system that wouldn’t be possible in a life system. This proved to be helpful especially during the testing stages. It also allowed me better control the layout of the system in the static case, as well as the control the repeatability of the
Directory: 16 K entries per node.
Disk: 7 (4 for data, 2 for database logs, 1 for OS)
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

12. Benchmark DSS queries based on TPC-H
Database was populated based on a TPC-H database with a scale factor of 1
Around 2.5 GB between table and indices
Data set of the queries ranged from 585 MB to 2.8 GB
Web interactions are based on TPC-W
DB2 was optimized for the simulated hardware running each type of workload
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

13. Performance for DSS workload
Static (global):
Generates a single cost function for a group of queries
Obtains the layout most suitable for all the queries
Static (local):
Uses a single query to produce a layout
Produces a very optimistic layout
Dynamic:
Starts with an optimized layout
Adapts layout as queries run on the system
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

14. Cost functions Inter-node data transferred Time estimate
Sum of all data blocks that are accessed from a remote disk
Time estimate
Time to access local/remote data
Time to operate on the data
Time:
Estimates the time to access the disk, transfer the data to memory and perform the operation
Then uses the maximum time of all the nodes
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

15. Quality of the solution: steps
If we reduce the temperature slowly, we can achieve a better schedule
This comes at the cost of extra time
Around 0.87 seconds of think-time for 50 steps
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

16. Summary
We phrase the data placement problem as a combinatorial optimization problem
We propose a technique that uses simulated annealing to generate an initial data layout based on the expected usage of the data
We extend the simulated annealing technique to reorganize the data at run-time
We take advantage of the locality of data references to improve the effectiveness of the reorganization
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement

17. Thank you For additional information
10/28/2004
Improving Server Performance on Transaction Processing Workloads by Enhanced Data Placement