Scribd
Upload
106 views

Vmware Esxi Optimization and Comparison With Baremetal Linux

This document compares the performance of a virtual machine (VM) hosted on VMware ESXi v6.7 U3 to a bare metal Linux system. Key findings include: 1) A single VM with 4 SCSI controllers and optimized configuration achieved 67-92% of the performance of the bare metal system for various workloads. 2) Using multiple VMs, overall performance approached that of the bare metal system, with two VMs achieving 80-95% of bare metal performance. 3) For most workloads tested, including sequential, random, and OLTP, there were negligible differences in performance between the optimized VM and bare metal configurations.

Uploaded by

smowaishyder
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
106 views

Vmware Esxi Optimization and Comparison With Baremetal Linux

This document compares the performance of a virtual machine (VM) hosted on VMware ESXi v6.7 U3 to a bare metal Linux system. Key findings include: 1) A single VM with 4 SCSI controllers and optimized configuration achieved 67-92% of the performance of the bare metal system for various workloads. 2) Using multiple VMs, overall performance approached that of the bare metal system, with two VMs achieving 80-95% of bare metal performance. 3) For most workloads tested, including sequential, random, and OLTP, there were negligible differences in performance between the optimized VM and bare metal configurations.

Uploaded by

smowaishyder
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 21

VMware ESXi v6.

7 U3
Optimization and
Comparison with Bare
Metal Linux
Ninh Nguyen
Center for Performance
Global Product & Solutions Enablement (GPSE)
September 2020

CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
Objective

• Optimize a Virtual Machine hosted by VMware ESXi v6.7 U3

• Compare performance results between Virtual Machine (VM) hosted by


ESXi and Bare Metal (BM) System
‒ 6-Corner workload – 4KB, 8KB, and 16KB Random, 256KB Sequential.
‒ OLTP workload

CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
Executive Summary (1/2)

• A VM with 4 SCSI controllers achieved the highest Random performance


(highest rate, lowest response time) compared to the VM with 1 or 2 SCSI.

• There are negligible differences in performance when using VMware


Paravirtual SCSI (PVSCSI), LSI Logic SAS, and LSI Logic Parallel SCSI
controller.

• There are insignificant differences in performance when tuning HBA


lpfc_lun_queue_depth and lpfc_link_speed parameter.

• Over allocating vCPUs degraded performance for Random workloads.

• Over allocating Memory did not impact in performance.

CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
Executive Summary (2/2)
• A single VM achieved about 67% RRCH, 77% RRCM, 89% RWCH, and 92%
RWCM of Bare metal system.
‒ VM achieved very much in the same range IOPS level when using 4KB and 8KB RRCH which
shows that this IOPS result is the maximum IOPS a single VM can achieve.
• Two VMs achieved about 95% RRCH, 80% RRCM, 95% RWCH and 90% RWCM
of Bare metal system.
‒ Random performance results saturated at 2 VMs which indicates that those results are the
maximum a ESXi with 4 x 32Gbps FC ports can achieve.
• There are negligible differences in performance between VM and BM for
Sequential workloads.
• There are negligible differences in performance between VM and BM when using
16KB Random workloads except for RRCM, VM achieved ~92% RRCM IOPS
compared to BM
• There are negligible differences in both IOPS and response time between VM and
BM for OLTP workloads.
CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
Test bed
Details
Microcode: SVOS 9.4 Q-Code (93-02-01-60/00)
Cache Size: 1024GB
Storage Components CPU: 56 Cores @ 2.3GHz (28 cores per MPU)
Drives: 48 x 3.8TB NVMe SSDs (SNR5A-R3R8NC), 4DBN with standard DKBNs
Ports: 32 x 32Gbps FC ports (4 CHB pairs, 4 ports per CHB)
VSP E990

Hitachi Vantara DS220


CPU: 2 x Intel(R) Xeon(R) Platinum 8176M CPU @ 2.10GHz
Memory: 256GB
2 x Servers
OS :
Bare Metal : RHEL 8.1
VM : OS RHEL 8.1 hosted by ESXi v6.7 U3
HBAs: 4 x Emulex LPe32000 PCIe Dual-Port 32G FC Adapters, 1 port per HBA was used
FC Switches Brocade G620 64port 32Gbps, FOS v8.0.2c, 16 ports were used.

• Hardware configuration of ESXi server and Bare Metal are identical


‒ 56 cores (112 vCPUs with Hyper-threading), 256GB, 4 x 32Gbps FC ports
• 24 x 308GB DP-Vols were mapped to BM using 4 x 32 Gbps FC ports
• 24 x 308GB DP-Vols were mapped to ESXi using 4 x 32 Gbps FC ports
‒ 24 RDMs were assigned to VM(s)
CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
VM Performance Tuning

© Hitachi Vantara LLC 2020. All Rights Reserved.


SCSI Controller Count

• A single VM configured with 4 SCSI controllers achieved the best performance for Random
workloads, highest rate and fastest response time
‒ 24 RDMs equally mapped across 2 or 4 SCSI controllers
‒ Maximum a VM supports: 4 SCSI controllers per VM, 15 RDMs per SCSI controller
• There are negligible differences in performance results for RWCM, SR and SW when using
one or more SCSI controllers on a VM.
CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
SCSI Controller Types

• There are insignificant differences in performance between VMware


Paravirtual SCSI (PVSCSI), LSI Logic SAS, and LSI Logic Parallel SCSI
controller
CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
Emulex parameters - lpfc

• There are insignificant differences in performance when tuning HBA lpfc_lun_queue_depth


and lpfc_link_speed parameter
‒ Default lpfc_link_speed = Auto
‒ Default lpfc_lun_queue_depth = 30
CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
vCPU Allocation – Random workloads

• Over Allocating vCPUs degraded performance for Random Workloads


‒ 960 KIOPS with 22 vCPUs vs. 470 KIOPS with 104 vCPUs, and 407 KIOPS with 110
vCPUs for RRCH
CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
vCPU Allocation – Sequential Workloads

• There are negligible differences in performance for Sequential workloads


when VM equipped with 16GB of RAM and 16 vCPUs or more vCPUs
CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
Memory Allocation – Random Workloads

• There are negligible differences in performance for Random workloads


when VM equipped 22 vCPUs and 8GB of RAM or more
CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
Memory Allocation – Sequential Workloads

• There are negligible differences in performance for Sequential workloads


when VM equipped with 22 vCPUs and 8GB of RAM or more
CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
ESXi v6.7 U3 VM vs. Bare
Metal Performance

© Hitachi Vantara LLC 2020. All Rights Reserved.


8KB Random

• A single VM achieved about 67% RRCH, 77% RRCM, 89% RWCH, and 92% RWCM of BM
• 2 x VMs achieved about 95% RRCH, 80% RRCM, 95% RWCH and 90% RWCM of BM
CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
256KB Sequential

• There are negligible differences in performance between VM and Bare


metal system for Sequential workloads
CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
4KB Random

• The maximum rates of a single VM could achieve compared to a Bare


metal are 66% RRCH, 72% RRCM, 85% RWCH and 97% RWCM
CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
16KB Random

• There are negligible differences in performance between VM and BM when using 16KB
Random workloads except for RRCM, VM achieved ~92% IOPS compared to BM
CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
OLTP

• There are negligible differences in both IOPS and response time


between VM and BM
CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara LLC 2020. All Rights Reserved.
Thank You

© Hitachi Vantara LLC 2020. All Rights Reserved.


CONFIDENTIAL – For use by Hitachi Vantara employees and other audiences under NDA only. © Hitachi Vantara Corporation 2019. All Rights Reserved.

You might also like