BMWG S. Kommu
Internet Draft VMware
Intended status: Informational J. Rapp
Expires: September 2017 VMware
March 13, 2017
Considerations for Benchmarking Network Virtualization Platforms
draft-skommu-bmwg-nvp-00.txt
Status of this Memo
This Internet-Draft is submitted in full conformance with the
provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF), its areas, and its working groups. Note that
other groups may also distribute working documents as Internet-
Drafts.
Internet-Drafts are draft documents valid for a maximum of six
months and may be updated, replaced, or obsoleted by other documents
at any time. It is inappropriate to use Internet-Drafts as
reference material or to cite them other than as "work in progress."
The list of current Internet-Drafts can be accessed at
http://www.ietf.org/ietf/1id-abstracts.txt
The list of Internet-Draft Shadow Directories can be accessed at
http://www.ietf.org/shadow.html
This Internet-Draft will expire on September 13, 2017.
Copyright Notice
Copyright (c) 2016 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents
(http://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
carefully, as they describe your rights and restrictions with
respect to this document. Code Components extracted from this
document must include Simplified BSD License text as described in
Kommu, Basler & Rapp Expires September 13, 2017 [Page 1]
Internet-Draft NVP Benchmarking Considerations March 2017
Section 4.e of the Trust Legal Provisions and are provided without
warranty as described in the Simplified BSD License.
Abstract
Current network benchmarking methodologies are focused on physical
networking components and do not consider the actual application
layer traffic patterns and hence do not reflect the traffic that
virtual networking components work with. The purpose of this
document is to distinguish and highlight benchmarking considerations
when testing and evaluating virtual networking components in the
data center.
Table of Contents
1. Introduction .................................................. 2!
2. Conventions used in this document ............................. 3!
3. Definitions ................................................... 4!
3.1. System Under Test (SUT) .................................. 4!
3.2. Network Virtualization Platform .......................... 4!
3.3. Micro-services ........................................... 6!
4. Scope ......................................................... 7!
4.1. Virtual Networking for Datacenter Applications ........... 7!
4.2. Interaction with Physical Devices ........................ 8!
5. Interaction with Physical Devices ............................. 8!
5.1. Server Architecture Considerations ...................... 11!
6. Security Considerations ...................................... 14!
7. IANA Considerations .......................................... 14!
8. Conclusions .................................................. 14!
9. References ................................................... 14!
9.1. Normative References .................................... 14!
9.2. Informative References .................................. 15!
Appendix A. Partial List of Parameters to Document .............. 16!
A.1. CPU ..................................................... 16!
A.2. Memory .................................................. 16!
A.3. NIC ..................................................... 16!
A.4. Hypervisor .............................................. 17!
A.5. Guest VM ................................................ 18!
A.6. Overlay Network Physical Fabric ......................... 18!
A.7. Gateway Network Physical Fabric ......................... 18!
1. Introduction
Datacenter virtualization that includes both compute and network
virtualization is growing rapidly as the industry continues to look
for ways to improve productivity, flexibility and at the same time
Kommu & Rapp Expires September 13, 2017 [Page 2]
Internet-Draft NVP Benchmarking Considerations March 2017
cut costs. Network virtualization, is comparatively new and
expected to grow tremendously similar to compute virtualization.
There are multiple vendors and solutions out in the market, each
with their own benchmarks to showcase why a particular solution is
better than another. Hence, the need for a vendor and product
agnostic way to benchmark multivendor solutions to help with
comparison and make informed decisions when it comes to selecting
the right network virtualization solution.
Applications traditionally have been segmented using VLANs and ACLs
between the VLANs. This model does not scale because of the 4K
scale limitations of VLANs. Overlays such as VXLAN were designed to
address the limitations of VLANs
With VXLAN, applications are segmented based on VXLAN encapsulation
(specifically the VNI field in the VXLAN header), which is similar
to VLAN ID in the 802.1Q VLAN tag, however without the 4K scale
limitations of VLANs. For a more detailed discussion on this
subject please refer RFC 7364 "Problem Statement: Overlays for
Network Virtualization".
VXLAN is just one of several Network Virtualization Overlays(NVO).
Some of the others include STT, Geneve and NVGRE. . STT and Geneve
have expanded on the capabilities of VXLAN. Please refer IETF's
nvo3 working group <
https://datatracker.ietf.org/wg/nvo3/documents/> for more
information.
Modern application architectures, such as Micro-services, are going
beyond the three tier app models such as web, app and db.
Benchmarks MUST consider whether the proposed solution is able to
scale up to the demands of such applications and not just a three-
tier architecture.
2. Conventions used in this document
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in RFC 2119 [RFC2119].
In this document, these words will appear with that interpretation
only when in ALL CAPS. Lower case uses of these words are not to be
interpreted as carrying significance described in RFC 2119.
Kommu & Rapp Expires September 13, 2017 [Page 3]
Internet-Draft NVP Benchmarking Considerations March 2017
3. Definitions
3.1. System Under Test (SUT)
Traditional hardware based networking devices generally use the
device under test (DUT) model of testing. In this model, apart from
any allowed configuration, the DUT is a black box from a testing
perspective. This method works for hardware based networking
devices since the device itself is not influenced by any other
components outside the DUT.
Virtual networking components cannot leverage DUT model of testing
as the DUT is not just the virtual device but includes the hardware
components that were used to host the virtual device
Hence SUT model MUST be used instead of the traditional device under
test
With SUT model, the virtual networking component along with all
software and hardware components that host the virtual networking
component MUST be considered as part of the SUT.
Virtual networking components may also work with higher level TCP
segments such as TSO. In contrast, all physical switches and
routers, including the ones that act as initiators for NVOs, work
with L2/L3 packets.
Please refer to section 5 Figure 1 for a visual representation of
System Under Test in the case of Intra-Host testing and section 5
Figure 2 for System Under Test in the case of Inter-Host testing
3.2. Network Virtualization Platform
This document does not focus on Network Function Virtualization.
Network Function Virtualization (NFV) focuses on being independent
of networking hardware while providing the same functionality. In
the case of NFV, traditional benchmarking methodologies recommended
by IETF may be used. Considerations for Benchmarking Virtual
Network Functions and Their Infrastructure IETF document addresses
benchmarking NFVs.
Typical NFV implementations emulate in software, the characteristics
and features of physical switches. They are similar to any physical
L2/L3 switch from the perspective of the packet size, which is
typically enforced based on the maximum transmission unit used.
Network Virtualization platforms on the other hand, are closer to
the application layer and are able to work with not only L2/L3
Kommu & Rapp Expires September 13, 2017 [Page 4]
Internet-Draft NVP Benchmarking Considerations March 2017
packets but also segments that leverage TCP optimizations such as
Large Segment Offload (LSO).
NVPs leverage TCP stack optimizations such as TCP Segmentation
Offload (TSO) and Large Receive Offload (LRO) that enables NVPs to
work with much larger payloads of up to 64K unlike their
counterparts such as NFVs.
Because of the difference in the payload, which translates into one
operation per 64K of payload in NVP verses ~40 operations for the
same amount of payload in NFV because of having to divide it to MTU
sized packets, results in considerable difference in performance
between NFV and NVP.
Please refer to figure 1 for a pictorial representation of this
primary difference between NPV and NFV for a 64K payload
segment/packet running on network set to 1500 bytes MTU.
Note: Payload sizes in figure 1 are approximates.
Kommu & Rapp Expires September 13, 2017 [Page 5]
Internet-Draft NVP Benchmarking Considerations March 2017
NPV (1 segment) NFV (40 packets)
Segment 1 Packet 1
+-------------------------+ +-------------------------+
| Headers | | Headers |
| +---------------------+ | | +---------------------+ |
| | Pay Load - upto 64K | | | | Pay Load < 1500 | |
| +---------------------+ | | +---------------------+ |
+-------------------------+ +-------------------------+
Packet 2
+-------------------------+
| Headers |
| +---------------------+ |
| | Pay Load < 1500 | |
| +---------------------+ |
+-------------------------+
.
.
.
.
Packet 40
+-------------------------+
| Headers |
| +---------------------+ |
| | Pay Load < 1500 | |
| +---------------------+ |
+-------------------------+
Figure 1 Payload NPV vs NFV
Hence, normal benchmarking methods are not relevant to the NVPs.
Instead, newer methods that take into account the built in
advantages of TCP provided optimizations MUST be used for testing
Network Virtualization Platforms.
3.3. Micro-services
Traditional monolithic application architectures such as the three
tier web, app and db architectures are hitting scale and deployment
limits for the modern use cases.
Micro-services make use of classic unix style of small app with
single responsibility.
Kommu & Rapp Expires September 13, 2017 [Page 6]
Internet-Draft NVP Benchmarking Considerations March 2017
These small apps are designed with the following characteristics:
Each application only does one thing - like unix tools
Small enough that you could rewrite instead of maintain
Embedded with a simple web container
Packaged as a single executable
Installed as daemons
Each of these applications are completely separate
Interact via uniform interface
REST (over HTTP/HTTPS) being the most common
With Micro-services architecture, a single web app of the three tier
application model could now have 100s of smaller apps dedicated to
do just one job.
These 100s of small one responsibility only services will MUST be
secured into their own segment - hence pushing the scale boundaries
of the overlay from both simple segmentation perspective and also
from a security perspective
4. Scope
This document does not address Network Function Virtualization has
been covered already by previous IETF documents
(https://datatracker.ietf.org/doc/draft-ietf-bmwg-virtual-
net/?include_text=1) the focus of this document is Network
Virtualization Platform where the network functions are an intrinsic
part of the hypervisor's TCP stack, working closer to the
application layer and leveraging performance optimizations such
TSO/RSS provided by the TCP stack and the underlying hardware.
4.1. Virtual Networking for Datacenter Applications
While virtualization is growing beyond the datacenter, this document
focuses on the virtual networking for east-west traffic within the
datacenter applications only. For example, in a three tier app such
web, app and db, this document focuses on the east-west traffic
between web and app. It does not address north-south web traffic
accessed from outside the datacenter. A future document would
address north-south traffic flows.
Kommu & Rapp Expires September 13, 2017 [Page 7]
Internet-Draft NVP Benchmarking Considerations March 2017
This document addresses scale requirements for modern application
architectures such as Micro-services to consider whether the
proposed solution is able to scale up to the demands of micro-
services application models that basically have 100s of small
services communicating on some standard ports such as http/https
using protocols such as REST
4.2. Interaction with Physical Devices
Virtual network components cannot be tested independent of other
components within the system. Example, unlike a physical router or
a firewall, where the tests can be focused directly solely on the
device, when testing a virtual router or firewall, multiple other
devices may become part of the system under test. Hence the
characteristics of these other traditional networking switches and
routers, LB, FW etc. MUST be considered.
! Hashing method used
! Over-subscription rate
! Throughput available
! Latency characteristics
5. Interaction with Physical Devices
In virtual environments, System Under Test (SUT) may often share
resources and reside on the same Physical hardware with other
components involved in the tests. Hence SUT MUST be clearly
defined. In this tests, a single hypervisor may host multiple
servers, switches, routers, firewalls etc.,
Intra host testing: Intra host testing helps in reducing the number
of components involved in a test. For example, intra host testing
would help focus on the System Under Test, logical switch and the
hardware that is running the hypervisor that hosts the logical
switch, and eliminate other components. Because of the nature of
virtual infrastructures and multiple elements being hosted on the
same physical infrastructure, influence from other components cannot
be completely ruled out. For example, unlike in physical
infrastructures, logical routing or distributed firewall MUST NOT be
benchmarked independent of logical switching. System Under Test
definition MUST include all components involved with that particular
test.
Kommu & Rapp Expires September 13, 2017 [Page 8]
Internet-Draft NVP Benchmarking Considerations March 2017
+---------------------------------------------------+
| System Under Test |
| +-----------------------------------------------+ |
| | Hyper-Visor | |
| | | |
| | +-------------+ | |
| | | NVP | | |
| | +-----+ | Switch/ | +-----+ | |
| | | VM1 |<------>| Router/ |<------>| VM2 | | |
| | +-----+ VW | Fire Wall/ | VW +-----+ | |
| | | etc., | | |
| | +-------------+ | |
| | Legend | |
| | VM: Virtual Machine | |
| | VW: Virtual Wire | |
| +------------------------_----------------------+ |
+---------------------------------------------------+
Figure 2 Intra-Host System Under Test
Inter host testing: Inter host testing helps in profiling the
underlying network interconnect performance. For example, when
testing Logical Switching, inter host testing would not only test
the logical switch component but also any other devices that are
part of the physical data center fabric that connects the two
hypervisors. System Under Test MUST be well defined to help with
repeatability of tests. System Under Test definition in the case of
inter host testing, MUST include all components, including the
underlying network fabric.
Figure 2 is a visual representation of system under test for inter-
host testing
Kommu & Rapp Expires September 13, 2017 [Page 9]
Internet-Draft NVP Benchmarking Considerations March 2017
+---------------------------------------------------+
| System Under Test |
| +-----------------------------------------------+ |
| | Hyper-Visor | |
| | +-------------+ | |
| | | NVP | | |
| | +-----+ | Switch/ | +-----+ | |
| | | VM1 |<------>| Router/ |<------>| VM2 | | |
| | +-----+ VW | Fire Wall/ | VW +-----+ | |
| | | etc., | | |
| | +-------------+ | |
| +------------------------_----------------------+ |
| ^ |
| | Network Cabling |
| v |
| +-----------------------------------------------+ |
| | Physical Networking Components | |
| | switches, routers, firewalls etc., | |
| +-----------------------------------------------+ |
| ^ |
| | Network Cabling |
| v |
| +-----------------------------------------------+ |
| | Hyper-Visor | |
| | +-------------+ | |
| | | NVP | | |
| | +-----+ | Switch/ | +-----+ | |
| | | VM1 |<------>| Router/ |<------>| VM2 | | |
| | +-----+ VW | Fire Wall/ | VW +-----+ | |
| | | etc., | | |
| | +-------------+ | |
| +------------------------_----------------------+ |
+---------------------------------------------------+
Legend
VM: Virtual Machine
VW: Virtual Wire
Figure 3 Inter-Host System Under Test
Virtual components have a direct dependency on the physical
infrastructure that is hosting these resources. Hardware
characteristics of the physical host impact the performance of the
virtual components. The components that are being tested and the
impact of the other hardware components within the hypervisor on the
performance of the SUT MUST be documented. Virtual component
performance is influenced by the physical hardware components within
the hypervisor. Access to various offloads such as TCP segmentation
Kommu & Rapp Expires September 13, 2017 [Page 10]
Internet-Draft NVP Benchmarking Considerations March 2017
offload, may have significant impact on performance. Firmware and
driver differences may also significantly impact results based on
whether the specific driver leverages any hardware level offloads
offered. Hence, all physical components of the physical server
running the hypervisor that hosts the virtual components MUST be
documented along with the firmware and driver versions of all the
components used to help ensure repeatability of test results. For
example, BIOS configuration of the server MUST be documented as some
of those changes are designed to improve performance. Please refer
to Appendix A for a partial list of parameters to document.
5.1. Server Architecture Considerations
When testing physical networking components, the approach taken is
to consider the device as a black-box. With virtual infrastructure,
this approach would no longer help as the virtual networking
components are an intrinsic part of the hypervisor they are running
on and are directly impacted by the server architecture used.
Server hardware components define the capabilities of the virtual
networking components. Hence, server architecture MUST be
documented in detail to help with repeatability of tests. And the
entire hardware and software components become the SUT.
5.1.1. Frame format/sizes within the Hypervisor
Maximum Transmission Unit (MTU) limits physical network component's
frame sizes. The most common max supported MTU for physical devices
is 9000. However, 1500 MTU is the standard. Physical network
testing and NFV uses these MTU sizes for testing. However, the
virtual networking components that live inside a hypervisor, may
work with much larger segments because of the availability of
hardware and software based offloads. Hence, the normal smaller
packets based testing is not relevant for performance testing of
virtual networking components. All the TCP related configuration
such as TSO size, number of RSS queues MUST be documented along with
any other physical NIC related configuration.
Virtual network components work closer to the application layer then
the physical networking components. Hence virtual network
components work with type and size of segments that are often not
the same type and size that the physical network works with. Hence,
testing virtual network components MUST be done with application
layer segments instead of the physical network layer packets.
5.1.2. Baseline testing with Logical Switch
Logical switch is often an intrinsic component of the test system
along with any other hardware and software components used for
Kommu & Rapp Expires September 13, 2017 [Page 11]
Internet-Draft NVP Benchmarking Considerations March 2017
testing. Also, other logical components cannot be tested
independent of the Logical Switch.
5.1.3. Repeatability
To ensure repeatability of the results, in the physical network
component testing, much care is taken to ensure the tests are
conducted with exactly the same parameters. Parameters such as MAC
addresses used etc.,
When testing NPV components with an application layer test tool,
there may be a number of components within the system that may not
be available to tune or to ensure they maintain a desired state.
Example: housekeeping functions of the underlying Operating System.
Hence, tests MUST be repeated a number of times and each test case
MUST be run for at least 2 minutes if test tool provides such an
option. Results SHOULD be derived from multiple test runs. Variance
between the tests SHOULD be documented.
5.1.4. Tunnel encap/decap outside the hypervisor
Logical network components may also have performance impact based on
the functionality available within the physical fabric. Physical
fabric that supports NVO encap/decap is one such case that has
considerable impact on the performance. Any such functionality that
exists on the physical fabric MUST be part of the test result
documentation to ensure repeatability of tests. In this case SUT
MUST include the physical fabric
5.1.5. SUT Hypervisor Profile
Physical networking equipment has well defined physical resource
characteristics such as type and number of ASICs/SoCs used, amount
of memory, type and number of processors etc., Virtual networking
components' performance is dependent on the physical hardware that
hosts the hypervisor. Hence the physical hardware usage, which is
part of SUT, for a given test MUST be documented. Example, CPU
usage when running logical router.
CPU usage changes based on the type of hardware available within the
physical server. For example, TCP Segmentation Offload greatly
reduces CPU usage by offloading the segmentation process to the NIC
card on the sender side. Receive side scaling offers similar
benefit on the receive side. Hence, availability and status of such
hardware MUST be documented along with actual CPU/Memory usage when
the virtual networking components have access to such offload
capable hardware.
Kommu & Rapp Expires September 13, 2017 [Page 12]
Internet-Draft NVP Benchmarking Considerations March 2017
Following is a partial list of components that MUST be documented -
both in terms of what's available and also what's used by the SUT -
o CPU - type, speed, available instruction sets (e.g. AES-NI)
o Memory - type, amount
o Storage - type, amount
o NIC Cards - type, number of ports, offloads available/used,
drivers, firmware (if applicable), HW revision
o Libraries such as DPDK if available and used
o Number and type of VMs used for testing and
o vCPUs
o RAM
o Storage
o Network Driver
o Any prioritization of VM resources
o Operating System type, version and kernel if applicable
o TCP Configuration Changes - if any
o MTU
o Test tool
o Workload type
o Protocol being tested
o Number of threads
o Version of tool
o For inter-hypervisor tests,
o Physical network devices that are part of the test
! Note: For inter-hypervisor tests, system under test
is no longer only the virtual component that is being
tested but the entire fabric that connects the
Kommu & Rapp Expires September 13, 2017 [Page 13]
Internet-Draft NVP Benchmarking Considerations March 2017
virtual components become part of the system under
test.
6. Security Considerations
Benchmarking activities as described in this memo are limited to
technology characterization of a Device Under Test/System Under Test
(DUT/SUT) using controlled stimuli in a laboratory environment, with
dedicated address space and the constraints specified in the
sections above.
The benchmarking network topology will be an independent test setup
and MUST NOT be connected to devices that may forward the test
traffic into a production network, or misroute traffic to the test
management network.
Further, benchmarking is performed on a "black-box" basis, relying
solely on measurements observable external to the DUT/SUT.
Special capabilities SHOULD NOT exist in the DUT/SUT specifically
for benchmarking purposes. Any implications for network security
arising from the DUT/SUT SHOULD be identical in the lab and in
production networks.
7. IANA Considerations
No IANA Action is requested at this time.
8. Conclusions
Network Virtualization Platforms, because of their proximity to the
application layer and since they can take advantage of TCP stack
optimizations, do not function on packets/sec basis. Hence,
traditional benchmarking methods, while still relevant for Network
Function Virtualization, are not designed to test Network
Virtualization Platforms. Also, advances in application
architectures such as micro-services, bring new challenges and need
benchmarking not just around throughput and latency but also around
scale. New benchmarking methods that are designed to take advantage
of the TCP optimizations or needed to accurately benchmark
performance of the Network Virtualization Platforms
9. References
9.1. Normative References
[RFC7364] T. Narten, E. Gray, D. Black, L. Fang, L. Kreeger, M.
Napierala, "Problem Statement: Overlays for Network Virtualization",
RFC 7364, October 2014, https://datatracker.ietf.org/doc/rfc7364/
Kommu & Rapp Expires September 13, 2017 [Page 14]
Internet-Draft NVP Benchmarking Considerations March 2017
[nv03] IETF, WG, Network Virtualization Overlays, <
https://datatracker.ietf.org/wg/nvo3/documents/>
9.2. Informative References
[1] A. Morton " Considerations for Benchmarking Virtual Network
Functions and Their Infrastructure", draft-ietf-bmwg-virtual-
net-03, < https://datatracker.ietf.org/doc/draft-ietf-bmwg-
virtual-net/?include_text=1>
Kommu & Rapp Expires September 13, 2017 [Page 15]
Internet-Draft NVP Benchmarking Considerations March 2017
Appendix A. Partial List of Parameters to Document
A.1. CPU
CPU Vendor
CPU Number
CPU Architecture
# of Sockets (CPUs)
# of Cores
Clock Speed (GHz)
Max Turbo Freq. (GHz)
Cache per CPU (MB)
# of Memory Channels
Chipset
Hyperthreading (BIOS Setting)
Power Management (BIOS Setting)
VT-d
A.2. Memory
Memory Speed (MHz)
DIMM Capacity (GB)
# of DIMMs
DIMM configuration
Total DRAM (GB)
A.3. NIC
Vendor
Model
Port Speed (Gbps)
Kommu & Rapp Expires September 13, 2017 [Page 16]
Internet-Draft NVP Benchmarking Considerations March 2017
Ports
PCIe Version
PCIe Lanes
Bonded
Bonding Driver
Kernel Module Name
Driver Version
VXLAN TSO Capable
VXLAN RSS Capable
Ring Buffer Size RX
Ring Buffer Size TX
A.4. Hypervisor
Hypervisor Name
Version/Build
Based on
Hotfixes/Patches
OVS Version/Build
IRQ balancing
vCPUs per VM
Modifications to HV
Modifications to HV TCP stack
Number of VMs
IP MTU
Flow control TX (send pause)
Flow control RX (honor pause)
Kommu & Rapp Expires September 13, 2017 [Page 17]
Internet-Draft NVP Benchmarking Considerations March 2017
Encapsulation Type
A.5. Guest VM
Guest OS & Version
Modifications to VM
IP MTU Guest VM (Bytes)
Test tool used
Number of NetPerf Instances
Total Number of Streams
Guest RAM (GB)
A.6. Overlay Network Physical Fabric
Vendor
Model
# and Type of Ports
Software Release
Interface Configuration
Interface/Ethernet MTU (Bytes)
Flow control TX (send pause)
Flow control RX (honor pause)
A.7. Gateway Network Physical Fabric
Vendor
Model
# and Type of Ports
Software Release
Interface Configuration
Interface/Ethernet MTU (Bytes)
Kommu & Rapp Expires September 13, 2017 [Page 18]
Internet-Draft NVP Benchmarking Considerations March 2017
Flow control TX (send pause)
Flow control RX (honor pause)
Kommu & Rapp Expires September 13, 2017 [Page 19]
Internet-Draft NVP Benchmarking Considerations March 2017
Authors' Addresses
Samuel Kommu
VMware
3401 Hillview Ave
Palo Alto, CA, 94304
Email: skommu@vmware.com
Jacob Rapp
VMware
3401 Hillview Ave
Palo Alto, CA, 94304
Email: jrapp@vmware.com
Kommu & Rapp Expires September 13, 2017 [Page 20]