Data Loading...
Windows Server 2012 NIC Optimization and Best Practices ... Flipbook PDF
Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN . A Dell EqualLogic Reference Architecture .
154 Views
100 Downloads
FLIP PDF 1.04MB
Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN A Dell EqualLogic Reference Architecture Dell Storage Engineering September 2013
Revisions Date
Description
July 2013
Initial release
September 2013
Updated configuration recommendations
THIS WHITE PAPER IS FOR INFORMATIONAL PURPOSES ONLY, AND MAY CONTAIN TYPOGRAPHICAL ERRORS AND TECHNICAL INACCURACIES. THE CONTENT IS PROVIDED AS IS, WITHOUT EXPRESS OR IMPLIED WARRANTIES OF ANY KIND. © 2013 Dell Inc. All Rights Reserved. Dell, the Dell logo, and other Dell names and marks are trademarks of Dell Inc. in the US and worldwide. All other trademarks mentioned herein are the property of their respective owners.
2
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
Table of contents Revisions ............................................................................................................................................................................................. 2 Acknowledgements .......................................................................................................................................................................... 5 Feedback ............................................................................................................................................................................................ 5 Executive summary .......................................................................................................................................................................... 5 1
Introduction ................................................................................................................................................................................ 6 1.1
Audience ........................................................................................................................................................................... 6
2
Technical overview .................................................................................................................................................................... 7
3
Test configurations and methodology .................................................................................................................................. 8 3.1
Simplified SAN.................................................................................................................................................................. 8
3.1.1 Base SAN configuration ................................................................................................................................................. 9 3.1.2 Congested SAN configuration .................................................................................................................................... 10 3.2
I/O execution and evaluation ...................................................................................................................................... 11
3.3
Statistical analysis ........................................................................................................................................................... 11
3.4
Test case sequence ....................................................................................................................................................... 11
3.4.1 Broadcom BCM57810 NDIS mode test case sequence ........................................................................................ 12 3.4.2 Broadcom BCM57810 iSOE mode test case sequence ......................................................................................... 14 3.4.3 Intel X520 test case sequence .................................................................................................................................... 14 4
Results and analysis ................................................................................................................................................................. 16 4.1
Baseline testing .............................................................................................................................................................. 16
4.1.1 Jumbo Frames ............................................................................................................................................................... 16 4.1.2 Flow control ................................................................................................................................................................... 16 4.1.3 Receive and Transmit buffers ..................................................................................................................................... 17 4.2
Other available performance tuning options ........................................................................................................... 17
4.2.1 Interrupt moderation .................................................................................................................................................... 17 4.2.2 Receive Side Scaling ..................................................................................................................................................... 18 4.2.3 Receive Segment Coalescing ..................................................................................................................................... 18 4.2.4 TCP Offload Engine ...................................................................................................................................................... 18 4.2.5 Large send offload ........................................................................................................................................................ 19 4.2.6 TCP checksum offload................................................................................................................................................. 19 4.2.7 TCP receive window auto-tuning .............................................................................................................................. 19 4.2.8 Delayed ACK algorithm ................................................................................................................................................ 19
3
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
4.2.9 Nagle’s algorithm .......................................................................................................................................................... 19 4.2.10
5
iSCSI Offload Engine ................................................................................................................................................ 20
4.3
Broadcom BCM57810 NDIS mode performance results....................................................................................... 21
4.4
Broadcom BCM57810 iSOE mode performance results ....................................................................................... 23
4.5
Intel X520 performance results .................................................................................................................................. 24
Best practice recommendations ........................................................................................................................................... 26 5.1
Broadcom BCM57810 NDIS mode recommended configuration ....................................................................... 26
5.2
Broadcom BCM57810 iSOE mode recommended configuration ....................................................................... 27
5.3
Intel X520 recommended configuration .................................................................................................................. 27
6
Conclusion ................................................................................................................................................................................ 28
A
Test configuration details ....................................................................................................................................................... 29
B
Network adapter and TCP stack configuration details ..................................................................................................... 30
C
B.1
Broadcom BCM57810 NDIS mode adapter options............................................................................................... 30
B.2
Configuring Broadcom BCM57810 adapter properties in NDIS mode ............................................................... 31
B.3
Broadcom BCM57810 iSOE mode adapter options ............................................................................................... 33
B.4
Configuring Broadcom BCM57810 adapter properties in iSOE mode ............................................................... 33
B.5
Intel X520 adapter options .......................................................................................................................................... 36
B.6
Configuring Intel X520 adapter properties ............................................................................................................... 37
B.7
Windows Server 2012 TCP stack options ................................................................................................................. 38
B.8
Configuring the Windows Server 2012 TCP stack .................................................................................................. 38
B.9
Disabling unused network adapter protocols......................................................................................................... 40
I/O parameters ......................................................................................................................................................................... 42
Additional resources .......................................................................................................................................................................44
4
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
Acknowledgements This best practice white paper was produced by the following members of the Dell Storage team: Engineering: Clay Cooper Technical Marketing: Guy Westbrook Editing: Camille Daily Additional contributors: Irek Zabka, Mike Kosacek, Kirt Gillum, Jaiwant Virk, Steve Williamson, and Tony Ansley
Feedback We encourage readers of this publication to provide feedback on the quality and usefulness of this information by sending an email to [email protected].
[email protected]
Executive summary This reference architecture explores the configuration options available for improving Dell™ EqualLogic™ ® ® PS Series SAN performance using the Broadcom BCM57810 or Intel X520 10 GbE network adapters and ® th Windows Server 2012 on a Dell™ PowerEdge™ 12 generation server. Recommended OS and NIC configurations are given based on the results of SAN performance testing.
5
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
1
Introduction Dell EqualLogic PS Series arrays provide a storage solution that delivers the benefits of consolidated networked storage in a self-managing iSCSI storage area network (SAN) that is affordable and easy to use, regardless of scale. In every iSCSI SAN environment, there are numerous configuration options at the storage host which can have an effect on overall SAN performance. These effects can vary based on the size and available bandwidth of the SAN, the host/storage port ratio, the amount of network congestion, the I/O workload profile and the overall utilization of system resources at the storage host. Furthermore, a particular setting might greatly improve SAN performance for a large block sequential workload while having an insignificant or slightly negative effect on a small block random workload. Another setting might improve SAN performance at the expense of host CPU utilization. Keeping the above in mind, this technical paper quantifies the effect on iSCSI performance, as measured in throughput and IOPS, of several configuration options within the Broadcom and Intel 10 GbE adapter properties and the Windows Server 2012 TCP stack during three common SAN workloads. It also takes into account the value of certain settings in congested network environments and when host CPU resources are at premium. From the results, recommended configurations for a EqualLogic PS Series SAN are given for each tested NIC type. In order to focus on the pure SAN performance benefits of the tested configuration options, Data Center Bridging (DCB) and Broadcom Switch Independent Partitioning, also known as NIC Partitioning (NPAR) were excluded from testing. Note: The performance data in this paper is presented relative to baseline configurations and is not intended to express maximum performance or benchmark results. Actual workload, host to array port ratios, and other factors may also affect performance.
1.1
Audience This technical white paper is for storage administrators, SAN system designers, storage consultants, or anyone who is tasked with configuring a host server as an iSCSI initiator to EqualLogic PS Series storage for use in a production SAN. It is assumed that all readers have experience in designing and/or administering a shared storage solution. Also, there are some assumptions made in terms of familiarity with all current Ethernet standards as defined by the Institute of Electrical and Electronic Engineers (IEEE) as well as TCP/IP and iSCSI standards as defined by the Internet Engineering Task Force (IETF).
6
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
2
Technical overview iSCSI SAN traffic takes place over an Ethernet network and consists of communication between PS Series array member network interfaces and the iSCSI initiator of storage hosts. The Broadcom BCM57810 ® NetXtreme II and the Intel X520 10GbE network adapters were used as the iSCSI initiators during this project. The Broadcom BCM57810 network adapter features iSCSI Offload Engine (iSOE) technology which offloads processing of the iSCSI stack to the network adapter. When using iSOE mode, the network adapter becomes a host bus adapter (HBA) and a host-based, software iSCSI initiator is not utilized. This is as opposed to non-iSOE mode in which the network adapter functions as a traditional NIC and works with a software iSCSI initiator. This paper refers to the non-iSOE mode of operation as NDIS mode. In Windows, NDIS refers to the Network Driver Interface Specification, a standard application programming interface (API) for NICs. The following three initiator modes of operation were tested: 1. Broadcom BCM57810 NDIS mode 2. Broadcom BCM57810 iSOE mode 3. Intel X520
Appendix B provides a detailed list of the tested configuration options and default values for each NIC type as well as for the Windows Server 2012 TCP stack.
7
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
3
Test configurations and methodology The following section addresses the reasoning behind SAN design decisions and details the SAN configurations. Performance testing methodology, test case sequence, and results analysis are also explained.
3.1
Simplified SAN Every effort was made to simplify and optimize the test configurations so that the performance effects of each option could be isolated. The following configuration and design elements helped to achieve this goal. • • • • • • •
Host Integration Toolkit v.4.5.0 with default settings All unused protocols disabled for each network adapter Eight volumes within a single storage pool, evenly distributed across array members An isolated SAN with no LAN traffic Load balancing (volume page movement) disabled on the array members DCB was not used The NIC bandwidth was not partitioned
Windows Server 2012 natively supports a new feature for NIC teaming called NIC Load Balancing and Failover (LBFO). Dell recommends using the Host Integration Tool kit (HIT Kit) and MPIO for any NIC that is connected to EqualLogic iSCSI storage. The use of LBFO is not recommended for NICs dedicated to SAN connectivity since it does not add any benefit over MPIO. Microsoft LBFO or other NIC vendor specific teaming can still be used for non-SAN connected interfaces. If DCB is enabled in a converged network, LBFO can also be used for NIC partitions that are not SAN connected. Load balancing is recommended for production environments because it can improve SAN performance over time by optimizing volume data location based on I/O patterns. It was disabled for performance testing to maintain consistent test results. It is enabled by default.
The following two SAN designs were chosen. See Appendix A for more detail about the hardware and software infrastructure.
8
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
3.1.1
Base SAN configuration The first SAN design chosen was a basic SAN with a redundant SAN fabric and an equal number of host and storage ports. Having a 1:1 host/storage port ratio is ideal from a bandwidth perspective. This helped to ensure that optimal I/O rates were achieved during lab testing. Figure 1 shows only the active ports of the PS Series array members.
Figure 1 • • • •
9
Physical diagram of the base SAN configuration
Two switches Two array members each with a single port A single host with two 10 GbE NIC ports A 1:1 storage/host port ratio
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
3.1.2
Congested SAN configuration The second SAN design was constructed to mimic a host port experiencing network congestion. In this SAN design a single host port was oversubscribed by four storage ports for a 4:1 storage/host port ratio. Since only one host port existed, the SAN fabric was reduced to a single switch. Figure 2 shows only the active ports of the PS Series array members. A non-redundant SAN fabric is not recommended for a production SAN environment.
Figure 2 • • • •
10
Physical diagram of the congested SAN configuration
One switch Four array members each with one port A single host with one 10 GbE NIC A 4:1 storage/host port ratio
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
3.2
I/O execution and evaluation Prior to each test run, the host was restarted to confirm configuration changes were in effect. After boot, the even distribution of iSCSI connections across host and storage ports and of active array member ports across SAN switches was confirmed. The following three vdbench workloads were run: • 8 KB transfer size, random I/O, 67% read • 256 KB transfer size, sequential I/O, 100% read • 256 KB transfer size, sequential I/O, 100% write
For every test case, each vdbench workload was run three times for twenty minute durations and the results were averaged. Vdbench IOPS results were used to evaluate 8K random workload performance. Vdbench throughput results were used to evaluate 256K sequential workload performance. Host and array member retransmission rates and CPU utilization were also examined. See Appendix C for a list of vdbench parameters.
3.3
Statistical analysis In order to ensure stable results, the relative standard deviation among the three performance results for each test case workload was calculated. When the relative standard deviation was greater than 1% for a given test case workload, that particular workload performance test was re-run.
3.4
Test case sequence The performance and effect on system resources of adapter and Windows Server 2012 TCP stack options were evaluated using the test cases listed below. Initially, tests were run to compare a totally default configuration to the baseline configuration chosen for testing. The baseline configuration included the following non-default settings: 1. Jumbo frames enabled 2. Flow control enabled (if not already enabled by default) 3. Maximum receive and transmit buffers (if applicable)
Each subsequent test case consisted of a single option being toggled from the default setting to show its effect relative to the baseline configuration defined above. In test cases where a technology option had an adapter and a corresponding OS setting, for example Receive Side Scaling, both settings were changed simultaneously prior to the test case execution.
11
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
3.4.1
Broadcom BCM57810 NDIS mode test case sequence The following tables show the test case sequence used to evaluate the effect of tested configuration options for the Broadcom BCM57810 in NDIS mode. Bold text indicates the changed value for each test scenario. Table 1
Baseline test case sequence for Broadcom BCM57810 NDIS mode
Test Frame case size
Flow Control
Rx / Tx buffers
Other adapter settings
Windows Server TCP stack setting
Comments
1
Standard
On
Default
Default
Default
Default configuration
2
Jumbo
On
Default
Default
Default
Jumbo performance effect
3
Jumbo
On
Maximum
Default
Default
Baseline configuration to evaluate all subsequent settings
Table 2 Test case sequence for Broadcom BCM57810 NDIS mode to evaluate other configuration options
12
Windows Server Other adapter TCP stack settings setting
Test Frame case size
Flow Control
Rx / Tx buffers
4
Jumbo
On
Maximum
Interrupt moderation disabled
Default
5
Jumbo
On
Maximum
Receive Side Scaling (RSS) disabled
RSS disabled
6
Jumbo
On
Maximum
RSS queues of Default 16
7
Jumbo
On
Maximum
Receive Side Coalescing (RSC) disabled
RSC disabled
8
Jumbo
On
Maximum
TCP Connection Offload enabled
Chimney enabled
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
Comments
RSS enabled by default at adapter and in Windows Server TCP stack.
Both settings required to activate TCP Offload Engine (TOE)
13
Test Frame case size
Flow Control
Rx / Tx buffers
Windows Server Other adapter TCP stack settings setting
9
Jumbo
On
Maximum
Default
Receive Window Autotuning disabled
10
Jumbo
On
Maximum
Default
Delayed ACK algorithm disabled
11
Jumbo
On
Maximum
Default
Nagle’s algorithm disabled
12
Jumbo
On
Maximum
Large Send Offload (LSO) disabled
Default
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
Comments
3.4.2
Broadcom BCM57810 iSOE mode test case sequence The following table shows the test case sequence used to evaluate the effect of tested configuration options for the Broadcom BCM57810 in iSOE mode. Bold text indicates the changed value for each test scenario. As can be seen in the table below, when in iSOE mode the Broadcom 57810 has a much more limited set of adapter options. Also, in iSOE mode the Windows Server 2012 TCP stack options have no effect since the entire iSCSI and TCP stack are offloaded to the adapter by design. Table 3
Test case sequence for Broadcom BCM57810 iSOE mode
Test Frame case size
Flow Control
Rx / Tx buffers
Other adapter settings
Windows Server TCP stack
Comments
1
Standard
Auto
N/A
N/A
N/A
Default configuration
2
Jumbo
Auto
N/A
N/A
N/A
Jumbo performance effect
3
Jumbo
On
N/A
N/A
N/A
Recommended settings
Bold text indicates the changed value for each test scenario.
3.4.3
Intel X520 test case sequence The following table shows the test case sequence used to evaluate the effect of tested configuration options for the Intel X520. Bold text indicates the changed value for each test scenario. Table 4
14
Baseline test case sequence for Intel X520
Test Frame case size
Flow Control
Rx / Tx buffers
Other adapter Windows Server Comments settings TCP stack
1
Standard
On
Default
Default
Default
Default configuration
2
Jumbo
On
Default
Default
Default
Jumbo performance effect
3
Jumbo
On
Maximum
Default
Default
Baseline configuration to evaluate all subsequent settings
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
Table 5
15
Test case sequence for Intel X520 to evaluate other configuration options Other adapter settings
Test Frame case size
Flow Control
Rx / Tx buffers
Windows Server Comments TCP stack
4
Jumbo
On
Maximum
Interrupt moderation disabled
Default
5
Jumbo
On
Maximum
Receive Side Scaling (RSS) disabled
RSS disabled
6
Jumbo
On
Maximum
RSS queues of 16
Default
RSS enabled by default at adapter and in Windows Server TCP stack.
7
Jumbo
On
Maximum
RSS profile of Default NUMA scaling
RSS enabled by default at adapter and in Windows Server TCP stack.
8
Jumbo
On
Maximum
Receive Side Coalescing (RSC) disabled
RSC disabled
9
Jumbo
On
Maximum
Default
Receive Window Autotuning disabled
10
Jumbo
On
Maximum
Default
Delayed ACK algorithm disabled
11
Jumbo
On
Maximum
Default
Nagle’s algorithm disabled
12
Jumbo
On
Maximum
Large Send Offload (LSO) disabled
Default
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
4
Results and analysis All test case performance results for each NIC mode and workload combination are presented in this section. For the sake of analysis, a 5% margin of error is assumed and only performance differences greater than this were acknowledged as significant. Based on the results, recommended configurations for each NIC mode will be given in Section 5.
4.1
Baseline testing The following non-default settings were used as a baseline configuration for further testing and evaluation.
4.1.1
Jumbo Frames Jumbo frames enable Ethernet frames with payloads greater than 1500 bytes. In situations where large packets make up the majority of traffic and additional latency can be tolerated, jumbo packets can reduce CPU utilization and improve wire efficiency. Figure 3 illustrates the dramatic effect that enabling jumbo frames can have on large block workloads. Significant throughput increases were observed for both read and write large block workloads on both the Broadcom and Intel network adapters in all supported operating modes (i.e., NDIS, iSOE). Intel X520 256K sequential read throughput -- Standard vs. Jumbo frames 700.0% 600.0% 500.0% 400.0% % relative
300.0% 200.0% 100.0% 0.0% Standard frames Figure 3
4.1.2
Jumbo frames
The performance effect on a 256K sequential read workload of enabling jumbo frames on an Intel X520
Flow control Flow control is a link-level mechanism that enables the adapter to respond to or to generate flow control (PAUSE) frames, helping to regulate network traffic. Flow control is enabled by default and is well-known to be of benefit in a congested network environment.
16
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
4.1.3
Receive and Transmit buffers Rx / Tx buffers are used by the adapter when copying data to memory. Increasing this value can enhance performance with only a slight increase in system memory utilization. Maximizing buffer allocation is particularly important on a server with heavy CPU utilization and can also be beneficial during times of network congestion.
4.2
Other available performance tuning options For both the network adapter and the Windows Server 2012 TCP stack, there are other options available which can have an effect on SAN performance under the right circumstances. Section 4.2 defines these options and discusses the results of the performance testing. It is important to understand that the performance results described below may not translate to all EqualLogic PS Series SAN environments. The material presented below identifies setting and workload combinations that have a clearly positive or negative impact on iSCSI SAN performance. It is recommended that each potential configuration change be evaluated in the environment prior to implementation. For more information on performance tuning options for the networking subsystem see Performance Tuning Guidelines for Windows Server 2012 at http://msdn.microsoft.com/en-us/library/windows/hardware/jj248719.aspx The tuning guidelines and recommendations in Performance Tuning Guidelines for Windows Server 2012 are for TCP/IP network interfaces in general and are not particular to iSCSI networking.
4.2.1
Interrupt moderation With interrupt moderation, a network adapter attempts to reduce the number of interrupts required by sending a single interrupt for multiple events rather than an interrupt for each event. While this can help lower CPU utilization it can also increase SAN latency. It is enabled by default on both the Broadcom and Intel adapters. Neither a significant performance enhancement nor degradation was observed when it was disabled.
17
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
4.2.2
Receive Side Scaling RSS balances incoming traffic across multiple CPU cores, up to one logical process per CPU core. This can improve the processing of receive-intensive workloads when the number of available logical processors outnumbers network adapters. It is enabled by default on both the Broadcom and Intel adapters and globally in Windows Server 2012. In addition to disabling RSS to test against the baseline (RSS enabled), two primary customizations to RSS were also tested: 1.
Increasing the number of RSS queues to the maximum – A larger number of queues increases network throughput at the expense of CPU utilization. 2. Changing the RSS load balancing profile from Closest Processor to NUMA Scaling – The Closest Processor setting dynamically balances the RSS load across the available processors while the NUMA Scaling dynamically assigns processors in a round robin across the available NUMA nodes.
Outright disabling of RSS reduced 256K sequential read throughput by 30% on the Intel adapter. No other significant performance effects were observed on other workloads or on the Broadcom adapter. Increasing the number of RSS queues or changing the RSS load balancing profile did not have a significant effect on any workload performance for either the Broadcom or Intel adapter.
4.2.3
Receive Segment Coalescing RSC enables the adapter to collect packets from a TCP/IP stream and combine them into larger packets, thus reducing the number of IP headers that must be processed. Like RSS, RSC is meant to benefit receive-intensive workloads. It is enabled by default on both the Broadcom and Intel adapters and globally in Windows Server 2012. When RSC was disabled, no significant performance effects were observed for any workload on either the Broadcom or Intel adapters.
4.2.4
TCP Offload Engine TOE offloads the processing of the entire TCP/IP stack to the network adapter and is available on the Broadcom 57810 network adapter. For TOE to be active, TCP connection offload must be enabled on the network adapter and chimney support must be enabled in Windows Server 2012, both of which are disabled by default. While TOE did not significantly improve performance in any of the three workloads, it did have a positive effect on storage array member retransmission and CPU utilization. In the congested SAN configuration during the 256K sequential read workload, the array member retransmission rate was reduced from .11% when disabled to zero when enabled. The CPU utilization decreased from ~6% to ~4%. See Figure 4 below for an illustration.
18
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
4.2.5
Large send offload LSO enables the adapter to offload the task of segmenting TCP messages into valid Ethernet frames. Because the adapter hardware is able to complete data segmentation much faster than operating system software, this feature may improve transmission performance. In addition, the adapter uses fewer CPU resources. It is enabled by default on both Broadcom and Intel adapters. Disabling LSO resulted in decreased performance in almost every workload.
4.2.6
TCP checksum offload Enables the adapter to verify received packet checksums and compute transmitted packet checksums. This can improve TCP performance and reduce CPU utilization and is enabled by default. With TCP checksum disabled, iSCSI connection instability and increased array packet retransmission were observed during testing.
4.2.7
TCP receive window auto-tuning TCP window auto tuning enables Windows Server 2012 to monitor TCP connection transmission rates and increase the size of the TCP receive window if necessary. It is enabled by default. Disabling auto-tuning reduced 256K sequential read workload throughput by 30% on both Broadcom and Intel adapters.
4.2.8
Delayed ACK algorithm The delayed ACK algorithm is a technique to improve TCP performance by combining multiple ACK responses into a single response. This algorithm has the potential to interact negatively with a TCP sender utilizing Nagle’s algorithm, since Nagle’s algorithm delays data transmission until a TCP ACK is received. Though disabling this algorithm had no effect on the performance of any tested workload, there are cases where disabling TCP delayed ACK for an iSCSI interface may improve performance. For example, poor read performance and unreliable failover have been observed during periods of network congestion on Microsoft iSCSI cluster nodes. In certain cases, disabling TCP delayed ACK on iSCSI interfaces might be recommended by Dell Enterprise Support.
4.2.9
Nagle’s algorithm Nagle’s algorithm is a technique to improve TCP performance by buffering output in the absence of an ACK response until a packet’s worth of output has been reached. This algorithm has the potential to interact negatively with a TCP receiver utilizing the delayed ACK algorithm, since the delayed ACK algorithm may delay sending an ACK under certain conditions up to 500 milliseconds. While disabling Nagle’s algorithm did not demonstrate an effect on the performance results of the sustained workloads, there may be cases where disabling Nagle’s algorithm on the storage host improves iSCSI SAN performance. Bursty SAN I/O or a high frequency of iSCSI commands can trigger ACK delays and increase write latency. As with disabling TCP delayed ACK, Dell Enterprise Support might recommend disabling Nagle’s algorithm on iSCSI interfaces.
19
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
4.2.10
iSCSI Offload Engine iSOE offloads the processing of the entire iSCSI stack to the network adapter and is available on the Broadcom 57810 network adapter. For iSOE to be active it must be enabled in the Broadcom Advanced Control Suite (BACS). By default it is disabled. Once the iSOE function is enabled for the adapter (and the NDIS function disabled), it disappears from the native Windows Server networking administration and monitoring tools and appears as a storage controller in Windows device manager. It must be configured and managed through BACS. In the base SAN configuration, iSOE increased the IOPS of the 8K random read/write workload by over 30% while increasing the throughput of the 256K sequential read workload by 10% relative to the baseline NDIS configuration. 256K sequential write throughput was unchanged. Like TOE, iSOE had a positive effect on storage array member retransmission and CPU utilization. In the congested SAN configuration during the 256K sequential read workload, the array member retransmission rate was reduced from .11% when disabled to zero when enabled. The CPU utilization decreased significantly from ~6% to ~1.5%. See Figure 4 below for an illustration. Broadcom 57810 256K sequential read CPU utilization and array retransmit % in congested network 7%
0.12%
6%
0.10%
5%
0.08%
4% 0.06% 3%
Array retransmit % 0.04%
2%
0.02%
1% 0%
0.00% NDIS
Figure 4
20
CPU utilization %
TOE
iSOE
The CPU utilization and array retransmit percentage of a Broadcom 57810 in NDIS, TOE, and iSOE mode while running the 256K sequential read workload in the congested SAN configuration.
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
4.3
Broadcom BCM57810 NDIS mode performance results Section 4.3 shows the performance results when using the different configuration options during the three tested workloads using the Broadcom BCM57810 network adapter in NDIS mode. Broadcom 57810 NDIS mode 8K random read/write IOPS--Baseline vs. additional settings 120.0% 100.0% 80.0% 60.0% 40.0% 20.0%
% relative
0.0%
Figure 5
21
Broadcom 57810 NDIS mode, 8K random 67% read workload, IOPS relative to the baseline configuration.
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
Broadcom 57810 NDIS mode 256K sequential read throughput-- Baseline vs. additional settings 120.0% 100.0% 80.0% 60.0% 40.0% 20.0%
% relative
0.0%
Figure 6
Broadcom 57810 NDIS mode, 256K sequential read workload, throughput relative to the baseline configuration.
Broadcom 57810 NDIS mode 256K sequential write throughput-- Baseline vs. additional settings 120.0% 100.0% 80.0% 60.0% 40.0% 20.0% % relative
0.0%
Figure 7
22
Broadcom 57810 NDIS mode, 256K sequential write workload, throughput relative to the baseline configuration.
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
4.4
Broadcom BCM57810 iSOE mode performance results Section 4.4 compares the performance results with iSOE mode to those of NDIS mode and NDIS mode with TOE during the three tested workloads using the Broadcom BCM57810 network adapter. Broadcom 57810 8K random read/write IOPS--NDIS vs. TOE vs iSOE 140% 120% 100% 80% % relative 60% 40% 20% 0% NDIS Figure 8
TOE
iSOE
Broadcom 57810 iSOE mode, 8K random 67% read workload, IOPS relative to the baseline NDIS configuration and to the baseline NDIS configuration with TOE enabled.
Broadcom 57810 256K sequential read throughput--NDIS vs. TOE vs iSOE 120% 100% 80% 60%
% relative
40% 20% 0% NDIS Figure 9
23
TOE
iSOE
Broadcom 57810 iSOE mode, 256K sequential read workload, throughput to the baseline NDIS configuration and to the baseline NDIS configuration with TOE enabled.
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
Broadcom 57810 256K sequential write throughput--NDIS vs. TOE vs iSOE 120% 100% 80% 60%
% relative
40% 20% 0% NDIS
TOE
iSOE
Figure 10 Broadcom 57810 iSOE mode, 256K sequential write workload, throughput relative to the baseline NDIS configuration and to the baseline NDIS configuration with TOE enabled.
4.5
Intel X520 performance results Section 4.5 shows the performance results with the different configuration options during the three tested workloads using the Intel X520 network adapter. Intel X520 8K random read/write IOPS-- Baseline vs. additional settings 120.0% 100.0% 80.0% 60.0% 40.0% 20.0% 0.0%
% relative
Figure 11 Intel X520, 8K random 67% read workload, IOPS relative to the baseline configuration.
24
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
Intel X520 256K sequential read throughput-- Baseline vs. additional settings 120.0% 100.0% 80.0% 60.0% 40.0% 20.0%
% relative
0.0%
Figure 12 Intel X520, 256K sequential read workload, throughput relative to the baseline configuration. Intel X520 256K sequential write throughput-- Baseline vs. additional settings 120.0% 100.0% 80.0% 60.0% 40.0% 20.0%
% relative
0.0%
Figure 13 Intel X520, 256K sequential write workload, throughput relative to the baseline configuration.
25
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
5
Best practice recommendations In this section, recommended configurations will be given based on the performance results and analysis detailed in Section 4. Non-default settings are recommended only when a compelling difference in performance for one or more workloads was observed, or when a setting is known to provide benefit during network congestion or heavy CPU utilization. Only the non-default settings are listed. For a complete list of tested options and default values as well as instructions on making configuration changes to the storage host, see Appendix B.
5.1
Broadcom BCM57810 NDIS mode recommended configuration Based on the performance results and analysis for each workload, the following NIC and OS configuration changes are recommended. Table 6 Setting
Default value
Recommended value
Jumbo packet
1514
9614
Receive buffers
0=Auto
3000
TCP Connection Offload
Disabled
Enabled
Transmit Buffers
0=Auto
5000
Table 7
26
Broadcom BCM57810 NDIS mode recommended adapter configuration
Broadcom BCM57810 NDIS mode recommended Windows Server 2012 TCP configuration
Setting
Default value
Recommended value
Chimney Offload State
Disabled
Enabled
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
5.2
Broadcom BCM57810 iSOE mode recommended configuration Based on the performance results and analysis for each workload, the following NIC configuration changes are recommended. Since all TCP functions for the SAN interfaces are offloaded in iSOE mode, the Windows Server 2012 TCP configuration has no effect and no changes are required. Table 8
5.3
Broadcom BCM57810 iSOE mode recommended adapter configuration
Setting
Default value
Recommended value
Flow control
Auto
Rx & Tx enabled
MTU
1500
9600
Intel X520 recommended configuration Based on the performance results and analysis for each workload, the following NIC configuration changes are recommended. Table 9
27
Intel X520 recommended adapter configuration
Setting
Default value
Recommended value
Jumbo packet
Disabled
Enabled
Receive Buffers
512
4096
Transmit Buffers
512
16384
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
6
Conclusion For an EqualLogic PS Series SAN, it is recommended that jumbo frames and flow control be enabled for both the Broadcom 57810 and the Intel X520. If not using the Broadcom iSCSI Offload Engine, receive and transmit buffers should also be maximized. When using the Broadcom BCM57810, TCP Offload Engine should be considered for its ability to decrease CPU utilization and also to lower retransmission rates in congested networks. The Broadcom BCM57810 iSCSI Offload Engine is another compelling option. Not only did it decrease retransmission rates in a congested network and lower CPU utilization by an even greater amount than did TOE, it also exhibited some performance benefits, particularly during the 8K random read/write workload. One thing to consider when using iSOE is the difference in administration procedures. Once iSOE is enabled, the network adapter disappears from the native Windows Server network management and monitoring tools and appears as a storage controller in Windows Server device manager. It must be configured and monitored in the Broadcom Advanced Control Suite.
28
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
A
29
Test configuration details Hardware
Description
Blade enclosure
Dell PowerEdge M1000e chassis: • CMC firmware: 4.30
Blade server
Dell PowerEdge M620 server: • Windows Server 2012 Datacenter • BIOS version: 1.6.0 • iDRAC firmware: 1.31.30 • (2) Intel® Xeon® E5-2650 • 64GB RAM • Dual Broadcom 57810S-k 10GbE CNA o NDIS driver: 7.4.23.0 o iSOE driver: 7.4.3.0 o Firmware: 7.4.8 • Dual Intel x520-k 10GbE CNA o Driver: 3.1.65.0 o Firmware: 14.0.12 • Dell EqualLogic Host Integration Toolkit v4.5.0
Blade I/O modules
(2) Dell 10Gb Ethernet Pass-through module
SAN switches
(2) Dell Force10 s4810 • Firmware: 8.3.12.1
SAN array members
(2) Dell EqualLogic PS6110XV • (2) 10GbE controllers • Firmware: 6.0.2
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
B
Network adapter and TCP stack configuration details This section provides more detail about the configuration options and default settings of the network adapter properties and the Windows Server 2012 TCP stack.
B.1
Broadcom BCM57810 NDIS mode adapter options The following table lists the tested adapter options for the Broadcom BCM57810 NetXtreme II 10 GigE NIC in NDIS mode along with the default value. Table 10
Broadcom BCM57810 NDIS mode adapter options
Setting
Default value
Flow control
Rx & Tx Enabled
Interrupt moderation
Enabled
Jumbo packet
1514
Large Send Offload V2
Enabled
Number of RSS queues
4
Receive buffers
0=Auto
Receive Side Scaling
Enabled
Recv Segment Coalescing
Enabled
TCP Connection Offload*
Disabled
TCP/UDP Checksum Offload
Rx & Tx Enabled
Transmit buffers
0=Auto
* To be enabled, this option must be enabled in the NIC adapter settings and Windows Server TCP Chimney offload must be enabled.
30
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
B.2
Configuring Broadcom BCM57810 adapter properties in NDIS mode Adapter properties for the Broadcom BCM57810 NDIS adapter can be set in the traditional Windows Server adapter properties dialog box in the Advanced tab or with the Broadcom Advanced Control Suite (BACS), a separate application from the native Windows management tools.
Figure 14 Windows 2012 adapter properties window for the Broadcom NDIS adapter
31
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
Figure 15 Configuring properties for the Broadcom NDIS adapter using the Broadcom Advanced Control Suite
32
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
B.3
Broadcom BCM57810 iSOE mode adapter options The following table lists the tested adapter options for the Broadcom BCM57810 NetXtreme II 10 GigE NIC in iSOE mode along with the default value. Table 11
Broadcom BCM57810 iSOE mode adapter options
Setting
Default value
Flow control
Auto
MTU*
1500
* Equivalent to jumbo packet option above. Once iSOE is enabled, only the Flow control and MTU options are available for configuration.
B.4
Configuring Broadcom BCM57810 adapter properties in iSOE mode Adapter properties for the Broadcom BCM57810 iSOE adapter must be set in BACS. After enabling iSOE mode with BACS, jumbo frames and flow control settings can be established.
33
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
Figure 16 Enabling iSOE mode in the Broadcom Advanced Control Suite
34
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
Figure 17 Configuring flow control in the Broadcom Advanced Control Suite
Figure 18 Configuring jumbo frames (MTU) in the Broadcom Advanced Control Suite
35
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
B.5
Intel X520 adapter options The following table lists the tested adapter options for the Intel X520 10 GigE NIC along with the default value. Table 12
36
Intel X520 adapter options
Setting
Default value
Interrupt Moderation
Enabled
Jumbo packet
Disabled
Large Send Offload V2
Enabled
Maximum Number of RSS Queues
8
Flow Control
Rx & Tx Enabled
Receive Buffers
512
Transmit Buffers
512
Receive Side Scaling
Enabled
Recv Segment Coalescing
Enabled
RSS load balancing profile
Closest Processor
IPv4 Checksum Offload
Enabled
TCP Checksum Offload
Enabled
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
B.6
Configuring Intel X520 adapter properties Adapter properties for the Intel X520 NDIS adapter can be set in the traditional Windows Server adapter properties window in the Advanced tab.
Figure 19 Windows 2012 adapter properties window for the Intel NDIS adapter
37
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
B.7
Windows Server 2012 TCP stack options The following table lists the tested TCP stack options for Windows Server 2012 along with the default value. Table 13
Windows Server 2012 TCP stack options
Setting
Default value
Receive-Side Scaling State*
Enabled
Chimney Offload State*
Disabled
Receive Window Auto-Tuning Level
Normal
Receive Segment Coalescing State*
Enabled
Delayed ACK algorithm
Enabled
Nagle's algorithm
Enabled
* There are network adapter options which correspond to these OS options. During test cases, the option’s test case value was achieved by changing the option setting at both the network adapter and the OS.
B.8
Configuring the Windows Server 2012 TCP stack Windows Server 2012 TCP stack options can be configuring using the netsh command in PowerShell.
Figure 20 Using the netsh command in Windows Server 2012 PowerShell.
38
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
To disabled delayed ACK and Nagle’s algorithm, create the following entries for each SAN interface subkey in the Windows Server 2012 registry: Subkey location: HKEY_LOCAL_MACHINE \ SYSTEM \ CurrentControlSet \ Services \ Tcpip \ Parameters \ Interfaces \ Entries: TcpAckFrequency TcpNoDelay Value type: REG_DWORD, number Value to disable: 1
39
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
B.9
Disabling unused network adapter protocols Unused protocols can be disabled in the Windows Server 2012 network adapter properties menu.
Figure 21 Disabling unused protocols in the Windows Server 2012 network adapter properties menu
40
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
Figure 22 Disabling NetBIOS in the Windows Server 2012 Advanced TCP/IP Settings for the network adapter
41
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
C
I/O parameters Vdbench SAN workloads were executed using the following parameters in the parameter file. Common parameters: hd=default hd=one,system=localhost
iSCSI volumes (random IO): sd=sd1,host=*,lun=\\.\PhysicalDrive1,size=102400m,threads=5 sd=sd2,host=*,lun=\\.\PhysicalDrive2,size=102400m,threads=5 sd=sd3,host=*,lun=\\.\PhysicalDrive3,size=102400m,threads=5 sd=sd4,host=*,lun=\\.\PhysicalDrive4,size=102400m,threads=5 sd=sd5,host=*,lun=\\.\PhysicalDrive5,size=102400m,threads=5 sd=sd6,host=*,lun=\\.\PhysicalDrive6,size=102400m,threads=5 sd=sd7,host=*,lun=\\.\PhysicalDrive7,size=102400m,threads=5 sd=sd8,host=*,lun=\\.\PhysicalDrive8,size=102400m,threads=5
iSCSI volumes (sequential IO on two arrays): sd=sd1,host=*,lun=\\.\PhysicalDrive1,size=30m,threads=5 sd=sd2,host=*,lun=\\.\PhysicalDrive2,size=30m,threads=5 sd=sd3,host=*,lun=\\.\PhysicalDrive3,size=30m,threads=5 sd=sd4,host=*,lun=\\.\PhysicalDrive4,size=30m,threads=5 sd=sd5,host=*,lun=\\.\PhysicalDrive5,size=30m,threads=5 sd=sd6,host=*,lun=\\.\PhysicalDrive6,size=30m,threads=5 sd=sd7,host=*,lun=\\.\PhysicalDrive7,size=30m,threads=5 sd=sd8,host=*,lun=\\.\PhysicalDrive8,size=30m,threads=5
iSCSI volumes (sequential IO on four arrays): sd=sd1,host=*,lun=\\.\PhysicalDrive1,size=45m,threads=5 sd=sd2,host=*,lun=\\.\PhysicalDrive2,size=45m,threads=5 sd=sd3,host=*,lun=\\.\PhysicalDrive3,size=45m,threads=5 sd=sd4,host=*,lun=\\.\PhysicalDrive4,size=45m,threads=5 sd=sd5,host=*,lun=\\.\PhysicalDrive5,size=45m,threads=5 sd=sd6,host=*,lun=\\.\PhysicalDrive6,size=45m,threads=5 sd=sd7,host=*,lun=\\.\PhysicalDrive7,size=45m,threads=5 sd=sd8,host=*,lun=\\.\PhysicalDrive8,size=45m,threads=5
8KB random 67% read workload:
42
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
wd=wd1,sd=(sd1-sd8),xfersize=8192,rdpct=100,skew=67 wd=wd2,sd=(sd1-sd8),xfersize=8192,rdpct=0,skew=33
256KB sequential read workload: wd=wd1,sd=(sd1-sd8),xfersize=262144,rdpct=100,seekpct=sequential 256KB sequential write workload: wd=wd1,sd=(sd1-sd8),xfersize=262144,rdpct=0,seekpct=sequential Runtime options: rd=rd1,wd=wd*,iorate=max,elapsed=1200,interval=5
43
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN
Additional resources Support.dell.com is focused on meeting your needs with proven services and support. DellTechCenter.com is an IT Community where you can connect with Dell Customers and Dell employees for the purpose of sharing knowledge, best practices, and information about Dell products and your installations. Referenced or recommended Dell publications: • EqualLogic Configuration Guide: http://en.community.dell.com/dell-groups/dtcmedia/m/mediagallery/19852516/download.aspx • EqualLogic Compatibility Matrix (ECM): http://en.community.dell.com/techcenter/storage/w/wiki/2661.equallogic-compatibilitymatrix.aspx • EqualLogic Switch Configuration Guides: http://en.community.dell.com/techcenter/storage/w/wiki/4250.switch-configuration-guides-bysis.aspx • The latest EqualLogic firmware updates and documentation (site requires a login): http://support.equallogic.com Force10 Switch documentation: http://www.force10networks.com/CSPortal20/KnowledgeBase/Documentation.aspx For EqualLogic best practices white papers, reference architectures, and sizing guidelines for enterprise applications and SANs, refer to Storage Infrastructure and Solutions Team Publications at: • http://dell.to/sM4hJT Other recommended publications • Performance Tuning Guidelines for Windows Server 2012: http://msdn.microsoft.com/en-us/library/windows/hardware/jj248719.aspx
44
BP1057 | Windows Server 2012 NIC Optimization and Best Practices with EqualLogic SAN