Database servers on chip multiprocessors: limitations and ... · Database servers on chip multiprocessors: limitations and opportunities N. Hardavellas I. Pandis R. Johnson N. Mancheril

Database servers on chip multiprocessors: limitations and opportunitiesN. HardavellasI. PandisR. Johnson

N. MancherilA. AilamakiB. Falsafi

Benjamin ReillySeptember 27, 2011

Tuesday, 27 September, 11

The fattened cache (and CPU)

Cache capacity = Cache latency

More data on hand, but higher cost to retrieve it

CPUs show similar trend in development: continually larger, and more complex


OVERVIEW

•Motivation

• Experiment design

• Results and observations

•What now?

• Summary and discussion


Dividing the CMPs

CMPs?Chip multiprocessors: several cores sharing on-chip resources (caches)

Vary in:•# of cores•# of hardware threads (“contexts”)•Execution order•Pipeline depth


The ‘Fat Camp’ (FC)

Key characteristics•Few, but powerful cores•Few (1-2) hardware contexts•OoO –– Out-of-Order execution•ILP –– Instruction-level parallelism

Core 0 Core 1Thread Context


Hiding data stalls: FC

OoOOut of Order Execution

ILPInstruction-level parallelism

op 0

op 1

op 2

Wait for... input

Compute op 1

op 2

earlier operations

a += bb += cd += e


The ‘Lean Camp’ (LC)

Key characteristics•Many, but weaker cores•Several (4+) hardware contexts•In-order execution (simpler)

Core 0

Core 1 Core 5

Core 2

Core 3

Core 4

Core 7

Core 6


Core 0Core 0Core 0Core 0Core 0Core 0Core 0Core 0Core 0Core 0

Hiding data stalls: LCHardware contexts interleaved in round-robin fashion, skipping contexts that are in data stalls.

Running

Idle (runnable)

Stalled (non-runnable)


(Un)saturated workloads

Workloads•DSS•OLTP

Number of requests• Saturated: work always available for each

hardware context•Unsaturated: work not always available


LC vs. FC Performance

LC has slower response timein unsaturated workloads

+12%(low ILP for FC)

+70%(high ILP for FC)



LC has higher throughputin saturated workloads

+70%(ILP not

significant for FC)



Observations:•FC spends 46-64% of execution on data stalls•At best (saturated workloads), LC spends 76-80% on computation


Data stall breakdown

Larger (and hence slower) caches are decreasingly optimal

Consider three components of data cache stalls:

1. Cache size


Data stall breakdownCPI contributions for :

OLTP DSS

L2 hit stalls responsible for an increasingly large portion of the CPI



2. Per-chip core integration

SMP CMP

Processing 4x 1-core 1x 4-core

L2 cache(s) 4MB / CPU16 MB shared

Fewer cores per chip = fewer L2 hits



3. On-chip core count

8 cores:• 9% superlinear increase in

throughput (for DSS)

16 cores:• 26% sublinear decrease

(OLTP)• Too much pressure on L2


How do we apply this?

1. Increase parallelism

• Divide!(more threads ⇒ more saturation)

• Pipeline/OLP (producer-consumer pairs)

• Partition input (not ideal; static and complex)


How do we apply this?

2. Improve data locality

• Reduce data stalls to help with unsaturated workloads

• Halt producers in favour of consumers

• Use cache-friendly algorithms

3.Use staged DBs

• Partition work by groups of relational operators


Summary & Discussion

1. LC typically performs better than FC

• LC is best under saturated workloads.

• Is there room for FC CMPs in DB applications?

2. L2 hits are a bottleneck

• Why were DBs ignored in HW design?

• How can we avoid incurring the cost of an L2 hit?


Database servers on chip multiprocessors: limitations and ... · Database servers on chip multiprocessors: limitations and opportunities N. Hardavellas I. Pandis R. Johnson N. Mancheril

Documents