Transcript Thin Servers with Smart Pipes

Thin Servers with Smart Pipes:
Designing SoC Accelerators for
Memcached
BOHUA KOU
JING GAO
Overview
 Motivation & Introduction to Memcached Servers
 Performance Analysis on Existing Servers
 Identify Inefficiencies and bottlenecks
 Thin Servers with Smart Pipes (TSSP) Architecture
 TSSP Performance Results
 Conclusion & Questions
2
Internet Service Workload
 Internet services and data grow rapidly
 Database servers along cannot maintain
performance with justified cost
 Other types of infrastructure needed to
allow the modern internet services to scale
3
Memcached Servers
 Distributed key-value stores
 Implemented using a hash table, keys are unique and each
key maps to only a single server at any time
 Easy to scale (connecting to additional servers is simple)
 Attractive because server interface is general
4
Workloads
 Memcached behavior can vary considerably based on the size and access frequency of the
objects it stores
 Designed five workloads that capture a wide range of behavior to explore bottlenecks and
inefficiencies with existing servers.
FixedSize:
fixed object size and uniform popularity distribution; small objects place the
greatest stress on Memcached performance
MicroBlog:
object size and popularity distribution on a sample of “tweets” collected from
Twitter
Wiki:
use the entire Wikipedia database, each object represents an individual article
in HTML format
ImgPreview:
sample thumbnail photos and associated view counts from Flickr
FriendFeed:
same as Microblog
5
Workload Characteristics
6
System Under Test
 Two different server systems
 High-end Xeon-based server vs low-power Atom-based server
 Three different classes of network interface cards (NIC)
 A total of 21 different system configurations
7
Simulation Results
(CPU Bottlenecks)
 On the left: requests-per-second for GET-requests to fixed-size
objects; Gap becomes larger when size is below 1KB
 On the right: performance bottleneck shifts from network to
CPU
 Xeon-class systems operate at
less than 1/8 of their theoretical
instruction throughput
 Atom-class systems 1/16 of their
theoretical instruction
throughput
8
Simulation Results
(Caching Bottlenecks)
 L1 Icache misses really bad
 L1 Dcache and L2 performs
moderately well
9
Simulation Results
(TLB and Branch Prediction Bottlenecks)
 TLB behaviors for Xeon comparable to recent characterization works
 Atom provides an insufficient ITLB which causes most of its instruction fetch stalls
 Branch misprediction rates is high for Atom due to its less capable branch predictor than Xeon
10
Simulation Results
(Impact of NIC quality)
11
Simulation Results
 even though the Atom is a low-power processor with a significantly lower peak power than the Xeon, its power
efficiency is worse
12
Overcoming Bottlenecks:
 Bottleneck: GET operation
 The most latency and throughput critical memcached task
 Up to 30:1 GET/SET ratio
 Microarchitecture Solution: Thin Servers with Smart Pipes (TSSP)
 Shift GET from core to a hardware pipeline
 Using UDP protocol to replace TCP
Thin Servers = Embedded class cores
Smart Pipes = Integrated, on-die NIC with nearby hardware
13
TSSP Architecture Overview
Atom Based
Low power core
Integrated NIC and Accelerator
 Integrated NIC
 Handle incoming requests
GET Request
 MMU
 Virtual address translation
All the other requests
 Memcached Accelerator
 Respond to GET request without software
interaction
 All other request types and memory
management are handled by software
14
NIC and Memcached Accelerator
Routes based on IP
address, port and protocol
(TCP)
Data
Center
Network
NIC
System MMU
SoC fabric
Process GET requests
Hardware
traversable
Hash table
Memcached Accelerator
15
Hardware and Software Data Structure
 Variable 1-255 bytes key -> 64-bit identifier key’
 Split MemCached’s lookup structure (hash table) to the
key and value storage (Slab-allocated memory)
 4 possible slots for a given hash
 Need update key’s last-access timestamp
16
Evaluation Results
 Implementation
 Memcached as an FPGA appliance
Table: TSSP energy efficiency comparison for FixedSize Workload
 Area
 8000 Look Up Tables, ~2% of the FPGA
 Power
 Estimated TSSP Total power = SoC Xilinx Zynq
platform + non-CPU components of Atom-based system
 Performance
 Measuring processing time for the memcached
applications, eg. FixedSize
 Expected greater improvements if implemented as an
ASIC rather than FPGA
17
Conclusion
Identify several system bottlenecks for high-performance and low-power CPUs
 Propose the TSSP design: low-power embedded class core + Memcached
accelerator
 Potential 6 ~ 16X improvement in energy efficiency over existing servers
18
Discussion
Evaluation of TSSP only used FixedSize workload for comparison, is it enough?
Analysis on TSSP power consumption is weak. Does the estimated power reliable?
TSSP also modifies the software, is the change easy to make?
19