Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud...
Transcript of Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud...
![Page 1: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/1.jpg)
Generalization vs Specialization in cloud computing
Gustavo Alonso
Systems Group
Department of Computer Science
ETH Zurich, Switzerland
![Page 2: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/2.jpg)
ETH Systems Group
www.systems.ethz.ch
![Page 3: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/3.jpg)
What is the cloud?
Primarily a change in business model driven by computing demand
-A shift from manufacturing to services
![Page 4: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/4.jpg)
What is big data?
Primarily a change in business model driven by the ability to process large data collections
-A shift toward customization and
personalization of “services” (from services to information & human
services)
![Page 5: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/5.jpg)
The case for general purpose =Economies of scale
![Page 6: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/6.jpg)
The case for specialization
Hardware
Big Data
New models
![Page 7: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/7.jpg)
Big data
• Social graphs
• Page rank
• Science data
• HPC
• Genomic
Hardware
• Multicore
• GPUs
• FPGAs
• RDMA
• Smart NICs
• Active memory
New Models
• Data center
• Appliances
• Cloud computing
![Page 8: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/8.jpg)
Big Data Example
Load
• High peak workloads
• High update rate spikes
SLAs
• Stringent response time requirements
• Extensibility over time
Features
• Predictability
• Accurate provisioning
![Page 9: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/9.jpg)
...
Aggregation
Layers
Replication
Groups
...
...
External Clients
Crescando
...
...
Split
Scan Thread
Scan Thread
Scan Thread
Scan Thread
Scan Thread
Merge
Input Queue
(Operations)
Input Queue
(Operations)
Output Queue
(Result Tuples)
Output Queue
(Result Tuples)
Unterbrunner et al, PVLDB’09
CRESCANDO
![Page 10: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/10.jpg)
Modern Hardware
![Page 11: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/11.jpg)
Slide courtesy of Torsten Hoefler (Systems Group, ETH Zürich)
![Page 12: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/12.jpg)
Example 1: Processors
Oracle’s SPARC M7 processor: “SQL in silicon” accelerators processing streams of data from memory:
Decompress, Scan, Select, Translate
From Oracle M7 documentation
![Page 13: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/13.jpg)
Accelerators to come
From Oracle M7 documentation
![Page 14: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/14.jpg)
Example 2: Architectures
Microsoft CatapultData center nodes with FPGA boards
FPGAs connected through their own network (Torus topology)
Search and deep learning applications
Microsoft Cypherbase
Texas Advanced Computing Center (TACC)
![Page 15: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/15.jpg)
Example 3: both!!
From INTEL
![Page 16: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/16.jpg)
The computer you know no longer exists (at least, the interesting ones)
![Page 17: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/17.jpg)
Analysis of MapReduce workloads:Microsoft: median job size < 14 GB
Yahoo: median job size < 12.5 GB
Facebook: 90% of jobs less than 100 GB
Fit in main memory
One server more efficient than a cluster
Adding memory to a big server better than using a cluster
Nobody ever got fired for using Hadoop on a ClusterA. Rowstron, D. Narayanan, A. Donnely, G. O’Shea, A. Douglas
HotCDP 2012, Bern, Switzerland
![Page 18: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/18.jpg)
Multicore vs cluster
Barthels et al., SIGMOD’15
2x1024M
The predominant architecture will not be one computer (multicore) but a networked set of processing elements, possibly with
limited or no cache coherency
![Page 19: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/19.jpg)
What is a computing node?
The form factor for a computing node in a data center is about to change
Disaggregation
(HP’s The Machine)
Knights Landing Die (INTEL)up to 72 cores2 ports Omni-PathBootable host
![Page 20: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/20.jpg)
05-Apr-16 20
Consensus in a Box (Istvan et al, NSD’16)
Xilinx VC709 Evaluation Board
SFP+
SFP+
SFP+
SFP+
DRAM (8GB)
FPGA
Networking Atomic
Broadcast
Replicated
key-value store
Reads
Writes
SW Clients /
Other nodes
Other nodes
Other nodes
TCP
Direct
Direct
![Page 21: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/21.jpg)
05-Apr-16 21
The system
X 12
10Gbps Switch
3 FPGA cluster
Clients • Drop-in replacement for memcached with Zookeeper’s replication
• Standard tools for benchmarking (libmemcached)• Simulating 100s of clients
Comm. over TCP/IP
Comm. over direct connections + Leader election
+ Recovery
![Page 22: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/22.jpg)
05-Apr-16 22
Latency of puts in a KVS
Consensus
15-35μs ~10μs
Memaslap(ixgbe)
TCP / 10Gbps Ethernet
~3μsDirect connections
![Page 23: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/23.jpg)
1000
10000
100000
1000000
10000000
1 10 100 1000
Thro
ugp
ut
(co
nse
nsu
s ro
un
ds/
s)
Consensus latency (us)
FPGA (Direct)
FPGA (TCP)
DARE* (Infiniband)
Libpaxos (TCP)
Etcd (TCP)
Zookeeper (TCP)
Specialized solutions
05-Apr-16 23
The benefit of specialization…
General purpose solutions
[1] Dragojevic et al. FaRM: Fast Remote Memory. In NSDI’14.[2] Poke et al. DARE: High-Performance State Machine Replication on RDMA Networks. In HPDC’15.*=We extrapolated from the 5 node setup for a 3 node setup.
10-100x
![Page 24: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/24.jpg)
Processing everywhere
![Page 25: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/25.jpg)
(Woods, VLDB’14; Istvan, SIGMOD’14)
IBEX
![Page 26: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/26.jpg)
A processor on the data path
![Page 27: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/27.jpg)
Sounds good?
Imagine the same at all levels:Smart storageOn the network switch (SDN like)On the network card (smart NIC)On the PCI express bus On the memory bus (active memory)
Every element in the system
(a computer rack)
will be a processing component
![Page 28: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/28.jpg)
In a cloud?
These developments will force rethinking many aspects of the cloud
scheduling
sharing and virtualization
platforms for storing and processing data
Ignoring the specilization through hardware not a good idea
performance / relevance
![Page 29: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/29.jpg)
This is the end …
![Page 30: Generalization vs Specialization in cloud computing...Generalization vs Specialization in cloud computing Gustavo Alonso Systems Group Department of Computer Science ETH Zurich, Switzerland](https://reader033.fdocuments.us/reader033/viewer/2022052717/5f045eb87e708231d40da439/html5/thumbnails/30.jpg)
The agenda ahead of us
• Very interesting times• Many opportunities driven by hardware
• Plenty of use cases justifying specialization
• Many challenges• Hardware changes affect the whole stack
• How to program heterogeneous architectures
• A new era in computer science• Architecture
• Data centers
• Economic pressures and models