SlideShare a Scribd company logo
1 of 14
Download to read offline
Dror Goldenberg | VP Software Architecture
DevelopEX 2015
Software at the Speed of Light
2November 11, 2015
We Live in a World of Data
Data Needs to be Accessible Always and in Real-Time
Data is Growing Exponentially
More DataMore ApplicationsMore Devices
3November 11, 2015
The Challenge - Efficient Data
Processing
StoreAnalyze
Enabling the Use of Data
4November 11, 2015 Figure: https://frankaxtell.bandcamp.com/track/the-speed-
of-light
100 Gigabit Ethernet
64B
Packe
t
1518B Packet
• 8M packets per second
• 123ns/packet
• 149M packets per
second
• 6.7ns/packet
5November 11, 2015 Figure: https://frankaxtell.bandcamp.com/track/the-speed-
of-light
100 Gigabit Ethernet
64B
Packe
t
1518B Packet
10.8
7.6
2.9
1.7
0
2
4
6
8
10
12
OPS/PACKET
Single 1518B Packet
Software Budget
0.59
0.42
0.16
0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
OPS/PACKET
Single 64B Packet
Software Budget
6November 11, 2015 Figure: https://frankaxtell.bandcamp.com/track/the-speed-
of-light
Storage Trends
10
100
1,000
10,000
100,000
1,000,000
10,000,000
100,000,000
Storage Access
Latency (ns)
Storage at 100GE
• 3M I/O per second (4KB
I/Os)
• 333us per I/O
28.9
20.35
7.8
4.7
0
5
10
15
20
25
30
35
OPS/PACKET
Storage I/O Software
Budget
7November 11, 2015
Addressing The Speed Challenges
Hardware Offloads
New Protocols
New APIs
Optimizations
8November 11, 2015
Hardware Offloads
•Scalability to multi CPU cores
•Reduce CPU cycles per operation
•Improve CPU efficiency (NUMA,
affinity, etc)
Single Root I/O Virtualization (SRIOV)
Legacy NIC
Hypervisor
vSwitch
Trusted
VM
V
M
V
M
V
M
PF
SW VNICSW VNICSW VNICSW VNICNIC
SRIOV I/O NIC
Hypervisor
eSwitch
Trusted
VM
V
M
V
M
V
M
PF VF VF VF
NICNICNICNIC
•Direct assignment of devices to VMs
•Improve VM I/O efficiency
•Avoid hypercalls
•Virtual switch offloads
•Enable advanced protocols
Stateless offloads
Kernel bypass
RDMA
T10/DIF Offload
Erasure Code
Offload
SRIOV
Examples
Collective Offloads
Overlay Network
Offloads
9November 11, 2015
New Protocols
•Reduce protocol stack overheads
•Add new functionality and
operations
NVMe Over Fabrics Block Storage Protocol
•Lightweight protocol stack
•Bypass storage stack layers
•Native mapping into storage devices
•Native multi queueing
•Low latency using RDMA (<10us
additional latency vs native)
•More HW offloads
App
Virtual
Filesystem
Block Layer
SCSI Mid Layer
iSCSI
RDMA Verbs
iSER
App
Virtual
Filesystem
Block Layer
NVMe over Fabrics
RDMA Verbs
RDMA over
Converged Ethernet
(RoCE)
Examples
NVMe over Fabrics
Overlay Networks
10November 11, 2015
New APIs
•Reduce protocol stack overheads
•Enable usermode protocol stacks
•Add new functionality and
operations
•Enable offloading
UCX
DPDK
RDMA
Examples
RDMA (RoCE and InfiniBand)
•Kernel bypass
•Message semantics
•Zero copy
•Protocol offload
•Batching operations
•Polling and interrupts
App
ULP/Middleware
RDMA Verbs Provider
RDMA I/F & Infrastructure
RDMA
Aware
App
11November 11, 2015
Software Optimizations
•Optimize overall system
performance
•Leverage offloads
•Amortize overheads
Batching
Polling/interrupt
Usermode Access
Avoid False Cache
Line Sharing
Data/Code
Compaction
Lock Avoidance
Examples
Scaling to multi
core
Affinitizing Flows
Polling vs Interrupt
Latency:
•Polling <1us
•Interrupt
•Bare Metal 5us
•VM 18us
X18 latency
False Cache Line Sharing
Transaction Per Second
& Shared counter on data path
•False sharing 2MTps
•Per core counter 3MTps
X1.5 throughput
Affinitizing Flows
Serving interrupts on the
transaction CPU core
•Before jittery ~50us/operation
•After ~5us/operation
X10 improvement
12November 11, 2015
Release N
Release N+1
Performance Bugs
Improved Performance
Better bandwidth for midsize
packets
Performance Bug
Reduced bandwidth on
large packets
Bandwidth(Gb/s)
Improving performance on
one workload can result in
hurting performance for
other workloads
13November 11, 2015
No Software Without Testing
•Performance benchmarks are critical
–Home grown
–Industry standard
•Performance regression
–Identify commits as soon as they break
performance
•Performance diagnostics
–Performance counters
•System performance tuning
14November 11, 2015 Figure: https://frankaxtell.bandcamp.com/track/the-speed-
of-light
Thank You!

More Related Content

What's hot

3 Ways to Automate App Deployments with NGINX
3 Ways to Automate App Deployments with NGINX3 Ways to Automate App Deployments with NGINX
3 Ways to Automate App Deployments with NGINXNGINX, Inc.
 
OpsStack--Integrated Operation Platform
OpsStack--Integrated Operation PlatformOpsStack--Integrated Operation Platform
OpsStack--Integrated Operation PlatformChinaNetCloud
 
NGINX Amplify: Monitoring NGINX with Advanced Filters and Custom Dashboards
NGINX Amplify: Monitoring NGINX with Advanced Filters and Custom DashboardsNGINX Amplify: Monitoring NGINX with Advanced Filters and Custom Dashboards
NGINX Amplify: Monitoring NGINX with Advanced Filters and Custom DashboardsNGINX, Inc.
 
ChinaNetCloud Online Lecture:Something About Tshark
ChinaNetCloud Online Lecture:Something About TsharkChinaNetCloud Online Lecture:Something About Tshark
ChinaNetCloud Online Lecture:Something About TsharkChinaNetCloud
 
Comprendre les différents types de Cloud
Comprendre les différents types de CloudComprendre les différents types de Cloud
Comprendre les différents types de CloudOVHcloud
 
Learning the basics of Apache NiFi for iot OSS Europe 2020
Learning the basics of Apache NiFi for iot OSS Europe 2020Learning the basics of Apache NiFi for iot OSS Europe 2020
Learning the basics of Apache NiFi for iot OSS Europe 2020Timothy Spann
 
Real time analytics with Netty, Storm, Kafka
Real time analytics with Netty, Storm, KafkaReal time analytics with Netty, Storm, Kafka
Real time analytics with Netty, Storm, KafkaTrieu Nguyen
 
Astricon 2012 - Redundancy and High Availability
Astricon 2012 - Redundancy and High AvailabilityAstricon 2012 - Redundancy and High Availability
Astricon 2012 - Redundancy and High Availabilityalipey
 
Best Practices for Monitoring DNS
Best Practices for Monitoring DNSBest Practices for Monitoring DNS
Best Practices for Monitoring DNSThousandEyes
 
Onboarding and Orchestrating High Performing Networking Software
Onboarding and Orchestrating High Performing Networking SoftwareOnboarding and Orchestrating High Performing Networking Software
Onboarding and Orchestrating High Performing Networking SoftwareCloudify Community
 
DevOps Fest 2020. Даніель Яворович. Data pipelines: building an efficient ins...
DevOps Fest 2020. Даніель Яворович. Data pipelines: building an efficient ins...DevOps Fest 2020. Даніель Яворович. Data pipelines: building an efficient ins...
DevOps Fest 2020. Даніель Яворович. Data pipelines: building an efficient ins...DevOps_Fest
 
FreeSWITCH as a Microservice
FreeSWITCH as a MicroserviceFreeSWITCH as a Microservice
FreeSWITCH as a MicroserviceEvan McGee
 
IT Press Tour #17 - OpenIO & Technology
IT Press Tour #17 - OpenIO & TechnologyIT Press Tour #17 - OpenIO & Technology
IT Press Tour #17 - OpenIO & TechnologyOpenIO Object Storage
 
Monitoring DNS Records and Servers
Monitoring DNS Records and ServersMonitoring DNS Records and Servers
Monitoring DNS Records and ServersThousandEyes
 
The Happy Marriage of Redis and Protobuf by Scott Haines of Twilio - Redis Da...
The Happy Marriage of Redis and Protobuf by Scott Haines of Twilio - Redis Da...The Happy Marriage of Redis and Protobuf by Scott Haines of Twilio - Redis Da...
The Happy Marriage of Redis and Protobuf by Scott Haines of Twilio - Redis Da...Redis Labs
 
IT Press Tour #19 Slides OpenIO June 2016
IT Press Tour #19 Slides OpenIO June 2016IT Press Tour #19 Slides OpenIO June 2016
IT Press Tour #19 Slides OpenIO June 2016OpenIO Object Storage
 
Robert Sander: CloudStack and Terraform
Robert Sander: CloudStack and TerraformRobert Sander: CloudStack and Terraform
Robert Sander: CloudStack and TerraformShapeBlue
 
OpenStack in the Enterprise
OpenStack in the EnterpriseOpenStack in the Enterprise
OpenStack in the EnterpriseTesora
 

What's hot (20)

3 Ways to Automate App Deployments with NGINX
3 Ways to Automate App Deployments with NGINX3 Ways to Automate App Deployments with NGINX
3 Ways to Automate App Deployments with NGINX
 
OpsStack--Integrated Operation Platform
OpsStack--Integrated Operation PlatformOpsStack--Integrated Operation Platform
OpsStack--Integrated Operation Platform
 
NGINX Amplify: Monitoring NGINX with Advanced Filters and Custom Dashboards
NGINX Amplify: Monitoring NGINX with Advanced Filters and Custom DashboardsNGINX Amplify: Monitoring NGINX with Advanced Filters and Custom Dashboards
NGINX Amplify: Monitoring NGINX with Advanced Filters and Custom Dashboards
 
ChinaNetCloud Online Lecture:Something About Tshark
ChinaNetCloud Online Lecture:Something About TsharkChinaNetCloud Online Lecture:Something About Tshark
ChinaNetCloud Online Lecture:Something About Tshark
 
Comprendre les différents types de Cloud
Comprendre les différents types de CloudComprendre les différents types de Cloud
Comprendre les différents types de Cloud
 
Learning the basics of Apache NiFi for iot OSS Europe 2020
Learning the basics of Apache NiFi for iot OSS Europe 2020Learning the basics of Apache NiFi for iot OSS Europe 2020
Learning the basics of Apache NiFi for iot OSS Europe 2020
 
Real time analytics with Netty, Storm, Kafka
Real time analytics with Netty, Storm, KafkaReal time analytics with Netty, Storm, Kafka
Real time analytics with Netty, Storm, Kafka
 
Astricon 2012 - Redundancy and High Availability
Astricon 2012 - Redundancy and High AvailabilityAstricon 2012 - Redundancy and High Availability
Astricon 2012 - Redundancy and High Availability
 
Best Practices for Monitoring DNS
Best Practices for Monitoring DNSBest Practices for Monitoring DNS
Best Practices for Monitoring DNS
 
ClueCon 2017
ClueCon 2017ClueCon 2017
ClueCon 2017
 
Onboarding and Orchestrating High Performing Networking Software
Onboarding and Orchestrating High Performing Networking SoftwareOnboarding and Orchestrating High Performing Networking Software
Onboarding and Orchestrating High Performing Networking Software
 
DevOps Fest 2020. Даніель Яворович. Data pipelines: building an efficient ins...
DevOps Fest 2020. Даніель Яворович. Data pipelines: building an efficient ins...DevOps Fest 2020. Даніель Яворович. Data pipelines: building an efficient ins...
DevOps Fest 2020. Даніель Яворович. Data pipelines: building an efficient ins...
 
FreeSWITCH as a Microservice
FreeSWITCH as a MicroserviceFreeSWITCH as a Microservice
FreeSWITCH as a Microservice
 
IT Press Tour #17 - OpenIO & Technology
IT Press Tour #17 - OpenIO & TechnologyIT Press Tour #17 - OpenIO & Technology
IT Press Tour #17 - OpenIO & Technology
 
OpenIO ServerLess Storage
OpenIO ServerLess StorageOpenIO ServerLess Storage
OpenIO ServerLess Storage
 
Monitoring DNS Records and Servers
Monitoring DNS Records and ServersMonitoring DNS Records and Servers
Monitoring DNS Records and Servers
 
The Happy Marriage of Redis and Protobuf by Scott Haines of Twilio - Redis Da...
The Happy Marriage of Redis and Protobuf by Scott Haines of Twilio - Redis Da...The Happy Marriage of Redis and Protobuf by Scott Haines of Twilio - Redis Da...
The Happy Marriage of Redis and Protobuf by Scott Haines of Twilio - Redis Da...
 
IT Press Tour #19 Slides OpenIO June 2016
IT Press Tour #19 Slides OpenIO June 2016IT Press Tour #19 Slides OpenIO June 2016
IT Press Tour #19 Slides OpenIO June 2016
 
Robert Sander: CloudStack and Terraform
Robert Sander: CloudStack and TerraformRobert Sander: CloudStack and Terraform
Robert Sander: CloudStack and Terraform
 
OpenStack in the Enterprise
OpenStack in the EnterpriseOpenStack in the Enterprise
OpenStack in the Enterprise
 

Viewers also liked

Lexical non equivalence- presentation
Lexical non equivalence- presentationLexical non equivalence- presentation
Lexical non equivalence- presentationEyhab Eddin
 
The problem of non equivalence
The problem of non equivalenceThe problem of non equivalence
The problem of non equivalenceEve_55
 
Non equivalene
Non equivaleneNon equivalene
Non equivalenesmallrat
 
3D Xpoint memory technology
3D Xpoint memory technology3D Xpoint memory technology
3D Xpoint memory technologyNITESH RAI
 
collocation
collocationcollocation
collocationsmallrat
 
Intel and DataStax: 3D XPoint and NVME Technology Cassandra Storage Comparison
Intel and DataStax: 3D XPoint and NVME Technology Cassandra Storage ComparisonIntel and DataStax: 3D XPoint and NVME Technology Cassandra Storage Comparison
Intel and DataStax: 3D XPoint and NVME Technology Cassandra Storage ComparisonDataStax Academy
 
Introduction to NVMe Over Fabrics-V3R
Introduction to NVMe Over Fabrics-V3RIntroduction to NVMe Over Fabrics-V3R
Introduction to NVMe Over Fabrics-V3RSimon Huang
 
Problems with non equivalence at word level
Problems with non equivalence at word levelProblems with non equivalence at word level
Problems with non equivalence at word levelThanh Phan Trung
 

Viewers also liked (10)

Lexical non equivalence- presentation
Lexical non equivalence- presentationLexical non equivalence- presentation
Lexical non equivalence- presentation
 
Olofiana.ppt copy
Olofiana.ppt   copyOlofiana.ppt   copy
Olofiana.ppt copy
 
The problem of non equivalence
The problem of non equivalenceThe problem of non equivalence
The problem of non equivalence
 
Non equivalene
Non equivaleneNon equivalene
Non equivalene
 
3D Xpoint memory technology
3D Xpoint memory technology3D Xpoint memory technology
3D Xpoint memory technology
 
Trans studies 3
Trans studies 3Trans studies 3
Trans studies 3
 
collocation
collocationcollocation
collocation
 
Intel and DataStax: 3D XPoint and NVME Technology Cassandra Storage Comparison
Intel and DataStax: 3D XPoint and NVME Technology Cassandra Storage ComparisonIntel and DataStax: 3D XPoint and NVME Technology Cassandra Storage Comparison
Intel and DataStax: 3D XPoint and NVME Technology Cassandra Storage Comparison
 
Introduction to NVMe Over Fabrics-V3R
Introduction to NVMe Over Fabrics-V3RIntroduction to NVMe Over Fabrics-V3R
Introduction to NVMe Over Fabrics-V3R
 
Problems with non equivalence at word level
Problems with non equivalence at word levelProblems with non equivalence at word level
Problems with non equivalence at word level
 

Similar to Dror goldenberg

Hail hydrate! from stream to lake using open source
Hail hydrate! from stream to lake using open sourceHail hydrate! from stream to lake using open source
Hail hydrate! from stream to lake using open sourceTimothy Spann
 
Qnap iei partners_day_2016 1108
Qnap iei partners_day_2016 1108Qnap iei partners_day_2016 1108
Qnap iei partners_day_2016 1108qnapivan
 
Music city data Hail Hydrate! from stream to lake
Music city data Hail Hydrate! from stream to lakeMusic city data Hail Hydrate! from stream to lake
Music city data Hail Hydrate! from stream to lakeTimothy Spann
 
CON6492 - Oracle Database Public Cloud Services v1 1
CON6492 - Oracle Database Public Cloud Services v1 1CON6492 - Oracle Database Public Cloud Services v1 1
CON6492 - Oracle Database Public Cloud Services v1 1David van Schalkwyk
 
Updates to Apache CloudStack and LINBIT SDS
Updates to Apache CloudStack and LINBIT SDSUpdates to Apache CloudStack and LINBIT SDS
Updates to Apache CloudStack and LINBIT SDSShapeBlue
 
Lagopus presentation on 14th Annual ON*VECTOR International Photonics Workshop
Lagopus presentation on 14th Annual ON*VECTOR International Photonics WorkshopLagopus presentation on 14th Annual ON*VECTOR International Photonics Workshop
Lagopus presentation on 14th Annual ON*VECTOR International Photonics WorkshopLagopus SDN/OpenFlow switch
 
IOT and System Platform From Concepts to Code
IOT and System Platform From Concepts to CodeIOT and System Platform From Concepts to Code
IOT and System Platform From Concepts to CodeAndy Robinson
 
Accelerating SDN Applications with Open Source Network Overlays
Accelerating SDN Applications with Open Source Network OverlaysAccelerating SDN Applications with Open Source Network Overlays
Accelerating SDN Applications with Open Source Network OverlaysCumulus Networks
 
Delivering High Performance Websites with NGINX
Delivering High Performance Websites with NGINXDelivering High Performance Websites with NGINX
Delivering High Performance Websites with NGINXNGINX, Inc.
 
G rpc talk with intel (3)
G rpc talk with intel (3)G rpc talk with intel (3)
G rpc talk with intel (3)Intel
 
Spark Streaming the Industrial IoT
Spark Streaming the Industrial IoTSpark Streaming the Industrial IoT
Spark Streaming the Industrial IoTJim Haughwout
 
Ceph Day SF 2015 - Deploying flash storage for Ceph without compromising perf...
Ceph Day SF 2015 - Deploying flash storage for Ceph without compromising perf...Ceph Day SF 2015 - Deploying flash storage for Ceph without compromising perf...
Ceph Day SF 2015 - Deploying flash storage for Ceph without compromising perf...Ceph Community
 
Ceph Day Amsterdam 2015 - Deploying flash storage for Ceph without compromisi...
Ceph Day Amsterdam 2015 - Deploying flash storage for Ceph without compromisi...Ceph Day Amsterdam 2015 - Deploying flash storage for Ceph without compromisi...
Ceph Day Amsterdam 2015 - Deploying flash storage for Ceph without compromisi...Ceph Community
 
Scale Your Load Balancer from 0 to 1 million TPS on Azure
Scale Your Load Balancer from 0 to 1 million TPS on AzureScale Your Load Balancer from 0 to 1 million TPS on Azure
Scale Your Load Balancer from 0 to 1 million TPS on AzureAvi Networks
 
QNAP NAS training 2016 Q3
QNAP NAS training 2016 Q3QNAP NAS training 2016 Q3
QNAP NAS training 2016 Q3qnapivan
 
How Tencent Applies Apache Pulsar to Apache InLong —— A Streaming Data Integr...
How Tencent Applies Apache Pulsar to Apache InLong —— A Streaming Data Integr...How Tencent Applies Apache Pulsar to Apache InLong —— A Streaming Data Integr...
How Tencent Applies Apache Pulsar to Apache InLong —— A Streaming Data Integr...StreamNative
 

Similar to Dror goldenberg (20)

IBM Aspera overview
IBM Aspera overview IBM Aspera overview
IBM Aspera overview
 
Hail hydrate! from stream to lake using open source
Hail hydrate! from stream to lake using open sourceHail hydrate! from stream to lake using open source
Hail hydrate! from stream to lake using open source
 
Qnap iei partners_day_2016 1108
Qnap iei partners_day_2016 1108Qnap iei partners_day_2016 1108
Qnap iei partners_day_2016 1108
 
Music city data Hail Hydrate! from stream to lake
Music city data Hail Hydrate! from stream to lakeMusic city data Hail Hydrate! from stream to lake
Music city data Hail Hydrate! from stream to lake
 
CON6492 - Oracle Database Public Cloud Services v1 1
CON6492 - Oracle Database Public Cloud Services v1 1CON6492 - Oracle Database Public Cloud Services v1 1
CON6492 - Oracle Database Public Cloud Services v1 1
 
Updates to Apache CloudStack and LINBIT SDS
Updates to Apache CloudStack and LINBIT SDSUpdates to Apache CloudStack and LINBIT SDS
Updates to Apache CloudStack and LINBIT SDS
 
Lagopus presentation on 14th Annual ON*VECTOR International Photonics Workshop
Lagopus presentation on 14th Annual ON*VECTOR International Photonics WorkshopLagopus presentation on 14th Annual ON*VECTOR International Photonics Workshop
Lagopus presentation on 14th Annual ON*VECTOR International Photonics Workshop
 
IOT and System Platform From Concepts to Code
IOT and System Platform From Concepts to CodeIOT and System Platform From Concepts to Code
IOT and System Platform From Concepts to Code
 
Accelerating SDN Applications with Open Source Network Overlays
Accelerating SDN Applications with Open Source Network OverlaysAccelerating SDN Applications with Open Source Network Overlays
Accelerating SDN Applications with Open Source Network Overlays
 
A new perspective on Network Visibility - RISK 2015
A new perspective on Network Visibility - RISK 2015A new perspective on Network Visibility - RISK 2015
A new perspective on Network Visibility - RISK 2015
 
Delivering High Performance Websites with NGINX
Delivering High Performance Websites with NGINXDelivering High Performance Websites with NGINX
Delivering High Performance Websites with NGINX
 
G rpc talk with intel (3)
G rpc talk with intel (3)G rpc talk with intel (3)
G rpc talk with intel (3)
 
RTI Connext 5.2.0
RTI Connext 5.2.0RTI Connext 5.2.0
RTI Connext 5.2.0
 
QNAP NAS Training 2016
QNAP NAS Training 2016QNAP NAS Training 2016
QNAP NAS Training 2016
 
Spark Streaming the Industrial IoT
Spark Streaming the Industrial IoTSpark Streaming the Industrial IoT
Spark Streaming the Industrial IoT
 
Ceph Day SF 2015 - Deploying flash storage for Ceph without compromising perf...
Ceph Day SF 2015 - Deploying flash storage for Ceph without compromising perf...Ceph Day SF 2015 - Deploying flash storage for Ceph without compromising perf...
Ceph Day SF 2015 - Deploying flash storage for Ceph without compromising perf...
 
Ceph Day Amsterdam 2015 - Deploying flash storage for Ceph without compromisi...
Ceph Day Amsterdam 2015 - Deploying flash storage for Ceph without compromisi...Ceph Day Amsterdam 2015 - Deploying flash storage for Ceph without compromisi...
Ceph Day Amsterdam 2015 - Deploying flash storage for Ceph without compromisi...
 
Scale Your Load Balancer from 0 to 1 million TPS on Azure
Scale Your Load Balancer from 0 to 1 million TPS on AzureScale Your Load Balancer from 0 to 1 million TPS on Azure
Scale Your Load Balancer from 0 to 1 million TPS on Azure
 
QNAP NAS training 2016 Q3
QNAP NAS training 2016 Q3QNAP NAS training 2016 Q3
QNAP NAS training 2016 Q3
 
How Tencent Applies Apache Pulsar to Apache InLong —— A Streaming Data Integr...
How Tencent Applies Apache Pulsar to Apache InLong —— A Streaming Data Integr...How Tencent Applies Apache Pulsar to Apache InLong —— A Streaming Data Integr...
How Tencent Applies Apache Pulsar to Apache InLong —— A Streaming Data Integr...
 

More from chiportal

Prof. Zhihua Wang, Tsinghua University, Beijing, China
Prof. Zhihua Wang, Tsinghua University, Beijing, China Prof. Zhihua Wang, Tsinghua University, Beijing, China
Prof. Zhihua Wang, Tsinghua University, Beijing, China chiportal
 
Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...
Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...
Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...chiportal
 
Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...
Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...
Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...chiportal
 
Prof. Uri Weiser,Technion
Prof. Uri Weiser,TechnionProf. Uri Weiser,Technion
Prof. Uri Weiser,Technionchiportal
 
Ken Liao, Senior Associate VP, Faraday
Ken Liao, Senior Associate VP, FaradayKen Liao, Senior Associate VP, Faraday
Ken Liao, Senior Associate VP, Faradaychiportal
 
Prof. Danny Raz, Director, Bell Labs Israel, Nokia
 Prof. Danny Raz, Director, Bell Labs Israel, Nokia  Prof. Danny Raz, Director, Bell Labs Israel, Nokia
Prof. Danny Raz, Director, Bell Labs Israel, Nokia chiportal
 
Marco Casale-Rossi, Product Mktg. Manager, Synopsys
Marco Casale-Rossi, Product Mktg. Manager, SynopsysMarco Casale-Rossi, Product Mktg. Manager, Synopsys
Marco Casale-Rossi, Product Mktg. Manager, Synopsyschiportal
 
Dr.Efraim Aharoni, ESD Leader, TowerJazz
Dr.Efraim Aharoni, ESD Leader, TowerJazzDr.Efraim Aharoni, ESD Leader, TowerJazz
Dr.Efraim Aharoni, ESD Leader, TowerJazzchiportal
 
Eddy Kvetny, System Engineering Group Leader, Intel
Eddy Kvetny, System Engineering Group Leader, IntelEddy Kvetny, System Engineering Group Leader, Intel
Eddy Kvetny, System Engineering Group Leader, Intelchiportal
 
Dr. John Bainbridge, Principal Application Architect, NetSpeed
 Dr. John Bainbridge, Principal Application Architect, NetSpeed  Dr. John Bainbridge, Principal Application Architect, NetSpeed
Dr. John Bainbridge, Principal Application Architect, NetSpeed chiportal
 
Xavier van Ruymbeke, App. Engineer, Arteris
Xavier van Ruymbeke, App. Engineer, ArterisXavier van Ruymbeke, App. Engineer, Arteris
Xavier van Ruymbeke, App. Engineer, Arterischiportal
 
Asi Lifshitz, VP R&D, Vtool
Asi Lifshitz, VP R&D, VtoolAsi Lifshitz, VP R&D, Vtool
Asi Lifshitz, VP R&D, Vtoolchiportal
 
Zvika Rozenshein,General Manager, EngineeringIQ
Zvika Rozenshein,General Manager, EngineeringIQZvika Rozenshein,General Manager, EngineeringIQ
Zvika Rozenshein,General Manager, EngineeringIQchiportal
 
Lewis Chu,Marketing Director,GUC
Lewis Chu,Marketing Director,GUC Lewis Chu,Marketing Director,GUC
Lewis Chu,Marketing Director,GUC chiportal
 
Kunal Varshney, VLSI Engineer, Open-Silicon
Kunal Varshney, VLSI Engineer, Open-SiliconKunal Varshney, VLSI Engineer, Open-Silicon
Kunal Varshney, VLSI Engineer, Open-Siliconchiportal
 
Gert Goossens,Sen. Director, ASIP Tools, Synopsys
Gert Goossens,Sen. Director, ASIP Tools, SynopsysGert Goossens,Sen. Director, ASIP Tools, Synopsys
Gert Goossens,Sen. Director, ASIP Tools, Synopsyschiportal
 
Tuvia Liran, Director of VLSI, Nano Retina
Tuvia Liran, Director of VLSI, Nano RetinaTuvia Liran, Director of VLSI, Nano Retina
Tuvia Liran, Director of VLSI, Nano Retinachiportal
 
Sagar Kadam, Lead Software Engineer, Open-Silicon
Sagar Kadam, Lead Software Engineer, Open-SiliconSagar Kadam, Lead Software Engineer, Open-Silicon
Sagar Kadam, Lead Software Engineer, Open-Siliconchiportal
 
Ronen Shtayer,Director of ASG Operations & PMO, NXP Semiconductor
Ronen Shtayer,Director of ASG Operations & PMO, NXP SemiconductorRonen Shtayer,Director of ASG Operations & PMO, NXP Semiconductor
Ronen Shtayer,Director of ASG Operations & PMO, NXP Semiconductorchiportal
 
Prof. Emanuel Cohen, Technion
Prof. Emanuel Cohen, TechnionProf. Emanuel Cohen, Technion
Prof. Emanuel Cohen, Technionchiportal
 

More from chiportal (20)

Prof. Zhihua Wang, Tsinghua University, Beijing, China
Prof. Zhihua Wang, Tsinghua University, Beijing, China Prof. Zhihua Wang, Tsinghua University, Beijing, China
Prof. Zhihua Wang, Tsinghua University, Beijing, China
 
Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...
Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...
Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...
 
Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...
Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...
Prof. Steve Furber, University of Manchester, Principal Designer of the BBC M...
 
Prof. Uri Weiser,Technion
Prof. Uri Weiser,TechnionProf. Uri Weiser,Technion
Prof. Uri Weiser,Technion
 
Ken Liao, Senior Associate VP, Faraday
Ken Liao, Senior Associate VP, FaradayKen Liao, Senior Associate VP, Faraday
Ken Liao, Senior Associate VP, Faraday
 
Prof. Danny Raz, Director, Bell Labs Israel, Nokia
 Prof. Danny Raz, Director, Bell Labs Israel, Nokia  Prof. Danny Raz, Director, Bell Labs Israel, Nokia
Prof. Danny Raz, Director, Bell Labs Israel, Nokia
 
Marco Casale-Rossi, Product Mktg. Manager, Synopsys
Marco Casale-Rossi, Product Mktg. Manager, SynopsysMarco Casale-Rossi, Product Mktg. Manager, Synopsys
Marco Casale-Rossi, Product Mktg. Manager, Synopsys
 
Dr.Efraim Aharoni, ESD Leader, TowerJazz
Dr.Efraim Aharoni, ESD Leader, TowerJazzDr.Efraim Aharoni, ESD Leader, TowerJazz
Dr.Efraim Aharoni, ESD Leader, TowerJazz
 
Eddy Kvetny, System Engineering Group Leader, Intel
Eddy Kvetny, System Engineering Group Leader, IntelEddy Kvetny, System Engineering Group Leader, Intel
Eddy Kvetny, System Engineering Group Leader, Intel
 
Dr. John Bainbridge, Principal Application Architect, NetSpeed
 Dr. John Bainbridge, Principal Application Architect, NetSpeed  Dr. John Bainbridge, Principal Application Architect, NetSpeed
Dr. John Bainbridge, Principal Application Architect, NetSpeed
 
Xavier van Ruymbeke, App. Engineer, Arteris
Xavier van Ruymbeke, App. Engineer, ArterisXavier van Ruymbeke, App. Engineer, Arteris
Xavier van Ruymbeke, App. Engineer, Arteris
 
Asi Lifshitz, VP R&D, Vtool
Asi Lifshitz, VP R&D, VtoolAsi Lifshitz, VP R&D, Vtool
Asi Lifshitz, VP R&D, Vtool
 
Zvika Rozenshein,General Manager, EngineeringIQ
Zvika Rozenshein,General Manager, EngineeringIQZvika Rozenshein,General Manager, EngineeringIQ
Zvika Rozenshein,General Manager, EngineeringIQ
 
Lewis Chu,Marketing Director,GUC
Lewis Chu,Marketing Director,GUC Lewis Chu,Marketing Director,GUC
Lewis Chu,Marketing Director,GUC
 
Kunal Varshney, VLSI Engineer, Open-Silicon
Kunal Varshney, VLSI Engineer, Open-SiliconKunal Varshney, VLSI Engineer, Open-Silicon
Kunal Varshney, VLSI Engineer, Open-Silicon
 
Gert Goossens,Sen. Director, ASIP Tools, Synopsys
Gert Goossens,Sen. Director, ASIP Tools, SynopsysGert Goossens,Sen. Director, ASIP Tools, Synopsys
Gert Goossens,Sen. Director, ASIP Tools, Synopsys
 
Tuvia Liran, Director of VLSI, Nano Retina
Tuvia Liran, Director of VLSI, Nano RetinaTuvia Liran, Director of VLSI, Nano Retina
Tuvia Liran, Director of VLSI, Nano Retina
 
Sagar Kadam, Lead Software Engineer, Open-Silicon
Sagar Kadam, Lead Software Engineer, Open-SiliconSagar Kadam, Lead Software Engineer, Open-Silicon
Sagar Kadam, Lead Software Engineer, Open-Silicon
 
Ronen Shtayer,Director of ASG Operations & PMO, NXP Semiconductor
Ronen Shtayer,Director of ASG Operations & PMO, NXP SemiconductorRonen Shtayer,Director of ASG Operations & PMO, NXP Semiconductor
Ronen Shtayer,Director of ASG Operations & PMO, NXP Semiconductor
 
Prof. Emanuel Cohen, Technion
Prof. Emanuel Cohen, TechnionProf. Emanuel Cohen, Technion
Prof. Emanuel Cohen, Technion
 

Dror goldenberg

  • 1. Dror Goldenberg | VP Software Architecture DevelopEX 2015 Software at the Speed of Light
  • 2. 2November 11, 2015 We Live in a World of Data Data Needs to be Accessible Always and in Real-Time Data is Growing Exponentially More DataMore ApplicationsMore Devices
  • 3. 3November 11, 2015 The Challenge - Efficient Data Processing StoreAnalyze Enabling the Use of Data
  • 4. 4November 11, 2015 Figure: https://frankaxtell.bandcamp.com/track/the-speed- of-light 100 Gigabit Ethernet 64B Packe t 1518B Packet • 8M packets per second • 123ns/packet • 149M packets per second • 6.7ns/packet
  • 5. 5November 11, 2015 Figure: https://frankaxtell.bandcamp.com/track/the-speed- of-light 100 Gigabit Ethernet 64B Packe t 1518B Packet 10.8 7.6 2.9 1.7 0 2 4 6 8 10 12 OPS/PACKET Single 1518B Packet Software Budget 0.59 0.42 0.16 0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 OPS/PACKET Single 64B Packet Software Budget
  • 6. 6November 11, 2015 Figure: https://frankaxtell.bandcamp.com/track/the-speed- of-light Storage Trends 10 100 1,000 10,000 100,000 1,000,000 10,000,000 100,000,000 Storage Access Latency (ns) Storage at 100GE • 3M I/O per second (4KB I/Os) • 333us per I/O 28.9 20.35 7.8 4.7 0 5 10 15 20 25 30 35 OPS/PACKET Storage I/O Software Budget
  • 7. 7November 11, 2015 Addressing The Speed Challenges Hardware Offloads New Protocols New APIs Optimizations
  • 8. 8November 11, 2015 Hardware Offloads •Scalability to multi CPU cores •Reduce CPU cycles per operation •Improve CPU efficiency (NUMA, affinity, etc) Single Root I/O Virtualization (SRIOV) Legacy NIC Hypervisor vSwitch Trusted VM V M V M V M PF SW VNICSW VNICSW VNICSW VNICNIC SRIOV I/O NIC Hypervisor eSwitch Trusted VM V M V M V M PF VF VF VF NICNICNICNIC •Direct assignment of devices to VMs •Improve VM I/O efficiency •Avoid hypercalls •Virtual switch offloads •Enable advanced protocols Stateless offloads Kernel bypass RDMA T10/DIF Offload Erasure Code Offload SRIOV Examples Collective Offloads Overlay Network Offloads
  • 9. 9November 11, 2015 New Protocols •Reduce protocol stack overheads •Add new functionality and operations NVMe Over Fabrics Block Storage Protocol •Lightweight protocol stack •Bypass storage stack layers •Native mapping into storage devices •Native multi queueing •Low latency using RDMA (<10us additional latency vs native) •More HW offloads App Virtual Filesystem Block Layer SCSI Mid Layer iSCSI RDMA Verbs iSER App Virtual Filesystem Block Layer NVMe over Fabrics RDMA Verbs RDMA over Converged Ethernet (RoCE) Examples NVMe over Fabrics Overlay Networks
  • 10. 10November 11, 2015 New APIs •Reduce protocol stack overheads •Enable usermode protocol stacks •Add new functionality and operations •Enable offloading UCX DPDK RDMA Examples RDMA (RoCE and InfiniBand) •Kernel bypass •Message semantics •Zero copy •Protocol offload •Batching operations •Polling and interrupts App ULP/Middleware RDMA Verbs Provider RDMA I/F & Infrastructure RDMA Aware App
  • 11. 11November 11, 2015 Software Optimizations •Optimize overall system performance •Leverage offloads •Amortize overheads Batching Polling/interrupt Usermode Access Avoid False Cache Line Sharing Data/Code Compaction Lock Avoidance Examples Scaling to multi core Affinitizing Flows Polling vs Interrupt Latency: •Polling <1us •Interrupt •Bare Metal 5us •VM 18us X18 latency False Cache Line Sharing Transaction Per Second & Shared counter on data path •False sharing 2MTps •Per core counter 3MTps X1.5 throughput Affinitizing Flows Serving interrupts on the transaction CPU core •Before jittery ~50us/operation •After ~5us/operation X10 improvement
  • 12. 12November 11, 2015 Release N Release N+1 Performance Bugs Improved Performance Better bandwidth for midsize packets Performance Bug Reduced bandwidth on large packets Bandwidth(Gb/s) Improving performance on one workload can result in hurting performance for other workloads
  • 13. 13November 11, 2015 No Software Without Testing •Performance benchmarks are critical –Home grown –Industry standard •Performance regression –Identify commits as soon as they break performance •Performance diagnostics –Performance counters •System performance tuning
  • 14. 14November 11, 2015 Figure: https://frankaxtell.bandcamp.com/track/the-speed- of-light Thank You!

Editor's Notes

  1. Assuming 3GHz CPU
  2. Ref: http://events.linuxfoundation.org/sites/events/files/slides/net_stack_challenges_100G_1.pdf http://www.slideshare.net/harryvanhaaren/the-7-deadly-sins-of-packet-processing-venky-venkatesan-and-bruce-richardson
  3. Assuming 3GHz CPU
  4. NUMA aware, false sharing, CPU+HW NUMA Cache line alignment Scaling to CPUs:RSS, Steering, Multithread/channel Interrupt affinity to core
  5. http://abovethecrowd.com/wp-content/uploads/2012/03/innovation.jpeg