Posts Tagged ‘EMC’

EMC Celerra Network Server Documentation

November 6th, 2010

EMC has updated their documentation library for the Celerra to version 6.0.  If you work with the Celerra or the UBER VSA, this is good reference documentation to have.  The updated Celerra documentation library on EMC’s Powerlink site is here: Celerra Network Server Documentation (User Edition) 6.0 A01.  The document library includes the following titles:

  • Celerra Network Server User Documents
    • Celerra CDMS Version 2.0 for NFS and CIFS
    • Celerra File Extension Filtering
    • Celerra Glossary
    • Celerra MirrorView/Synchronous Setup on CLARiiON Backends
    • Celerra Network Server Command Reference Manual
    • Celerra Network Server Error Messages Guide
    • Celerra Network Server Parameters Guide
    • Celerra Network Server System Operations
    • Celerra Security Configuration Guide
    • Celerra SMI-S Provider Programmer’s Guide
    • Configuring and Managing CIFS on Celerra
    • Configuring and Managing Celerra Network High Availability
    • Configuring and Managing Celerra Networking
    • Configuring Celerra Events and Notifications
    • Configuring Celerra Naming Services
    • Configuring Celerra Time Services
    • Configuring Celerra User Mapping
    • Configuring iSCSI Targets on Celerra
    • Configuring NDMP Backups on Celerra
    • Configuring NDMP Backups to Disk on Celerra
    • Configuring NFS on Celerra
    • Configuring Standbys on Celerra
    • Configuring Virtual Data Movers for Celerra
    • Controlling Access to Celerra System Objects
    • Getting Started with Celerra Startup Assistant
    • Installing Celerra iSCSI Host Components
    • Installing Celerra Management Applications
    • Managing Celerra for a Multiprotocol Environment
    • Managing Celerra Statistics
    • Managing Celerra Volumes and File Systems Manually
    • Managing Celerra Volumes and File Systems with Automatic Volume Management
    • Problem Resolution Roadmap for Celerra
    • Using Celerra AntiVirus Agent
    • Using Celerra Data Deduplication
    • Using Celerra Event Enabler
    • Using Celerra Event Publishing Agent
    • Using Celerra FileMover
    • Using Celerra Replicator (V2)
    • Using EMC Utilities for the CIFS Environment
    • Using File-Level Retention on Celerra
    • Using FTP on Celerra
    • Using International Character Sets with Celerra
    • Using MirrorView Synchronous with Celerra for Disaster Recovery
    • Using MPFS on Celerra
    • Using Multi-Protocol Directories with Celerra
    • Using NTMigrate with Celerra
    • Using ntxmap for Celerra CIFS User Mapping
    • Using Quotas on Celerra
    • Using SnapSure on Celerra
    • Using SNMPv3 on Celerra
    • Using SRDF/A with Celerra
    • Using SRDF/S with Celerra for Disaster Recovery
    • Using TFTP on Celerra Network Server
    • Using the Celerra nas_stig Utility
    • Using the Celerra server_archive Utility
    • Using TimeFinder/FS, NearCopy, and FarCopy with Celerra
    • Using Windows Administrative Tools with Celerra
    • Using Wizards to Configure Celerra
  • NS-120
    • Celerra NS-120 System (Single Blade) Installation Guide
    • Celerra NS-120 System (Dual Blade) Installation Guide
  • NS-480
    • Celerra NS-480 System (Dual Blade) Installation Guide
    • Celerra NS-480 System (Four Blade) Installation Guide
  • NS20
    • Celerra NS20 Read Me First
    • Setting Up the EMC Celerra NS20 System
    • Celerra NS21 Cabling Guide
    • Celerra NS21FC Cabling Guide
    • Celerra NS22 Cabling Guide
    • Celerra NS22FC Cabling Guide
    • Celerra NS20 System (Single Blade) Installation Guide
    • Celerra NS20 System (Single Blade with FC Option Enabled) Installation Guide
    • Celerra NS20 System (Dual Blade) Installation Guide
    • Celerra NS20 System (Dual Blade with FC Option Enabled) Installation Guide
  • NX4
    • Celerra NX4 System Single Blade Installation Guide
    • Celerra NX4 System Dual Blade Installation Guide
  • Regulatory Documents
    • C-RoHS HS/TS Substance Concentration Chart Technical Note

If you’re looking for more Celerra documentation, check out the Celerra Network Server General Reference page.

Configure VMware ESX(i) Round Robin on EMC Storage

February 4th, 2010

I recently set out to enable VMware ESX(i) 4 Round Robin load balancing with EMC Celerra (CLARiiON) fibre channel storage.  Before I get to the details of how I did it, let me preface this discussion with a bit about how I interpret Celerra storage architecture. 

The Celerra is built on CLARiiON fibre channel storage and as such, it leverages the benefits and successes CLARiiON has built over the years.  I believe most CLARiiON’s are, by default, active/passive arrays from VMware’s perspective.  Maybe more accurately stated, all controllers are active, however, each controller has sole ownership of a LUN or set of LUNs.  If a host wants access to a LUN, it is preferable to go through the owning controller (the preferred path).  Attempts to access a LUN through any other controller than the owning controller will result in a “Trespass” in EMC speak.  A Trespass is shift in LUN ownership from one controller to another in order to service an I/O request from a fabric host.  When I first saw Trespasses in Navisphere, I was alarmed.  I soon learned that they aren’t all that bad in moderation.  EMC reports that a Trespass occurs EXTREMELY quickly and in almost all cases will not cause problems.  However, as with any array which adopts the LUN ownership model, stacking up enough I/O requests which force a race condition between controllers for LUN access, will cause a condition known as thrashing.   Thrashing causes storage latency and queuing as controllers play tug of war for LUN access.  This is why it is important for ESX hosts, which share LUN access, to consistently access LUNs via the same controller path.  

As I said, the LUN ownership model above is the “out-of-box” configuration for the Celerra, also known as Failover Mode 1 in EMC Navisphere.  The LUN path going through the owning controller will be the Active path from a VMware perspective.  Other paths will be Standby.  This is true for both MRU and Fixed path selection policies.  What I needed to know was how to enable Round Robin path selection in VMware.  Choosing Round Robin in the vSphere Client is easy enough, however, there’s more to it than that because the Celerra is still operating in Failover Mode 1 where I/O can only go through the owning controller. 

So the first step in this process is to read the CLARiiON/VMware Applied Technology Guide which says I need to change the Failover Mode of the Celerra from 1 to 4 using Navisphere (FLARE release 28 version 04.28.000.5.704 or later may be required).  A value of 4 tells the CLARiiON to switch to the ALUA (Asymmetric Logical Unit Access or Active/Active) mode.  In this mode, the controller/LUN ownership model still exists, however, instead of transferring ownership of the LUN to the other controller with a Trespass, LUN access is allowed through the non-owning controller.  The I/O is passed by the non-owning controller to the owning controller via the backplane and then to the LUN.  In this configuration, both controllers are Active and can be used to access a LUN without causing ownership contention or thrashing.  It’s worth mentioning right now that although both controllers are active, the Celerra will report to ESX the owning controller as the optimal path, and the non-owning controller as the non-optimal path.  This information will be key a little later on.  Each ESX host needs to be configured for Failover Mode 4 in Navisphere.  The easiest way to do this is to run the Failover Setup Wizard.  Repeat the process for each ESX host.  One problem I ran into here is that after making the configuration change, each host and HBA still showed a Failover Mode of 1 in the Navisphere GUI.  It was as if the Failover Setup Wizard steps were not persisting.  I failed to accept this so I installed the Navisphere CLI and verified each host with the following command: 

naviseccli -h <SPA_IP_ADDRESS> port -list –all

Output showed that Failover Mode 4 was configured:

Information about each HBA:
HBA UID:                 20:00:00:00:C9:8F:C8:C4:10:00:00:00:C9:8F:C8:C4
Server Name:             lando.boche.mcse
Server IP Address:       192.168.110.5
HBA Model Description:
HBA Vendor Description:  VMware ESX 4.0.0
HBA Device Driver Name:
Information about each port of this HBA:�
    SP Name:               SP A
    SP Port ID:            2
    HBA Devicename:        naa.50060160c4602f4a50060160c4602f4a
    Trusted:               NO
    Logged In:             YES
    Source ID:             66560
    Defined:               YES
    Initiator Type:           3
    StorageGroup Name:     DL385_G2
    ArrayCommPath:         1
    Failover mode:         4
    Unit serial number:    Array

Unfortunately, the CLARiiON/VMware Applied Technology Guide didn’t give me the remaining information I needed to actually get ALUA and Round Robin working.  So I turned to social networking and my circle of VMware and EMC storage experts on Twitter.  They put me on to the fact that I needed to configure SATP for VMW_SATP_ALUA_CX, something I wasn’t familiar with yet. 

So the next step is a multistep procedure to configure the Pluggable Storage Architecture on the ESX hosts.  More specifically, SATP (Storage Array Type Plugin) and the PSP (Path Selection Plugin), in that order. Duncan Epping provides a good foundation for PSA which can be learned here.

Configuring the SATP tells the PSA what type of array we’re using, and more accurately, what failover mode the array is running.  In this case, I needed to configure the SATP for each LUN to VMW_SATP_ALUA_CX which is the EMC CLARiiON (CX series) running in ALUA mode (active/active failover mode 4).  The command to do this must be issued on each ESX host in the cluster for each active/active LUN and is as follows: 

#set SATP
esxcli nmp satp setconfig –config VMW_SATP_ALUA_CX –device naa.50060160c4602f4a50060160c4602f4a
esxcli nmp satp setconfig –config VMW_SATP_ALUA_CX –device naa.60060160ec242700be1a7ec7a208df11
esxcli nmp satp setconfig –config VMW_SATP_ALUA_CX –device naa.60060160ec242700bf1a7ec7a208df11
esxcli nmp satp setconfig –config VMW_SATP_ALUA_CX –device naa.60060160ec2427001cac9740a308df11
esxcli nmp satp setconfig –config VMW_SATP_ALUA_CX –device naa.60060160ec2427001dac9740a308df11

The devices you see above can be found in the vSphere Client when looking at the HBA devices discovered.  You can also find devices with the following command on the ESX Service Console: 

esxcli nmp device list 

I found that changing the SATP requires a host reboot for the change to take effect (thank you Scott Lowe).  After the host is rebooted, the same command used above should reflect that the SATP has been set correctly: 

esxcli nmp device list 

Results in: 

naa.60060160ec2427001dac9740a308df11
    Device Display Name: DGC Fibre Channel Disk (naa.60060160ec2427001dac9740a308df11)
    Storage Array Type: VMW_SATP_ALUA_CX
    Storage Array Type Device Config: {navireg=on, ipfilter=on}{implicit_support=on;explicit_ow=on;alua_followover=on;{TPG_id=1,TPG_state=ANO}{TPG_id=2,TPG_state=AO}}
    Path Selection Policy: VMW_PSP_FIXED
    Path Selection Policy Device Config: {policy=rr,iops=1000,bytes=10485760,useANO=0;lastPat=0,numBytesPending=0}
    Working Paths: vmhba1:C0:T0:L61 

Once the SATP is set, it is time to configure the PSP for each LUN to Round Robin.  You can do this via the vSphere Client, or you can issue the commands at the Service Console: 

#set PSP per device
esxcli nmp psp setconfig –config VMW_PSP_RR –device naa.60060160ec242700be1a7ec7a208df11
esxcli nmp psp setconfig –config VMW_PSP_RR –device naa.60060160ec242700bf1a7ec7a208df11
esxcli nmp psp setconfig –config VMW_PSP_RR –device naa.60060160ec2427001cac9740a308df11
esxcli nmp psp setconfig –config VMW_PSP_RR –device naa.60060160ec2427001dac9740a308df11 

#set PSP for device
esxcli nmp device setpolicy –psp VMW_PSP_RR –device naa.50060160c4602f4a50060160c4602f4a
esxcli nmp device setpolicy –psp VMW_PSP_RR –device naa.60060160ec242700be1a7ec7a208df11
esxcli nmp device setpolicy –psp VMW_PSP_RR –device naa.60060160ec242700bf1a7ec7a208df11
esxcli nmp device setpolicy –psp VMW_PSP_RR –device naa.60060160ec2427001cac9740a308df11
esxcli nmp device setpolicy –psp VMW_PSP_RR –device naa.60060160ec2427001dac9740a308df11 

Once again, running the command: 

esxcli nmp device list 

Now results in: 

naa.60060160ec2427001dac9740a308df11
    Device Display Name: DGC Fibre Channel Disk (naa.60060160ec2427001dac9740a308df11)
    Storage Array Type: VMW_SATP_ALUA_CX
    Storage Array Type Device Config: {navireg=on, ipfilter=on}{implicit_support=on;explicit_ow=on;alua_followover=on;{TPG_id=1,TPG_state=ANO}{TPG_id=2,TPG_state=AO}}
    Path Selection Policy: VMW_PSP_RR
    Path Selection Policy Device Config: {policy=rr,iops=1000,bytes=10485760,useANO=0;lastPat=0,numBytesPending=0}
    Working Paths: vmhba1:C0:T0:L61 

Notice the Path Selection Policy has now changed to Round Robin. 

I’m good to go, right?  Wrong.  I struggled with this last bit for a while.  Using ESXTOP and IOMETER, I could see that I/O was still only going down one path instead of two.  Then I remembered something Duncan Epping had said to me in an earlier conversation a few days ago.  He mentioned something about the array reporting optimal and non-optimal paths to the PSA.  I printed out a copy of the Storage Path and Storage Plugin Management with esxcli document from VMware and took it to lunch with me.  The answer was buried on page 88.  The nmp roundrobin setting useANO is configured by default to 0 which means unoptimized paths reported by the array will not be included in Round Robin path selection unless optimized paths become unavailable.  Remember I said early on that unoptimized and optimized paths reported by the array would be a key piece of information.  We can see this in action by looking at the device list above.  The very last line shows working paths, and only one path is listed for Round Robin use – the optimized path reported by the array.  The fix here is to issue the following command, again on each host for all LUNs in the configuration: 

#use non-optimal paths for Round Robin
esxcli nmp roundrobin setconfig –useANO 1 –device naa.50060160c4602f4a50060160c4602f4a
esxcli nmp roundrobin setconfig –useANO 1 –device naa.60060160ec242700be1a7ec7a208df11
esxcli nmp roundrobin setconfig –useANO 1 –device naa.60060160ec242700bf1a7ec7a208df11
esxcli nmp roundrobin setconfig –useANO 1 –device naa.60060160ec2427001cac9740a308df11
esxcli nmp roundrobin setconfig –useANO 1 –device naa.60060160ec2427001dac9740a308df11

Once again, running the command: 

esxcli nmp device list 

Now results in: 

naa.60060160ec2427001dac9740a308df11
    Device Display Name: DGC Fibre Channel Disk (naa.60060160ec2427001dac9740a308df11)
    Storage Array Type: VMW_SATP_ALUA_CX
    Storage Array Type Device Config: {navireg=on, ipfilter=on}{implicit_support=on;explicit_support=on;explicit_allow=on;alua_followover=on;{TPG_id=1,TPG_state=ANO}
TPG_id=2,TPG_state=AO}}
    Path Selection Policy: VMW_PSP_RR
    Path Selection Policy Device Config: {policy=rr,iops=1000,bytes=10485760,useANO=1;lastPathIndex=1: NumIOsPending=0,numBytesPending=0}
    Working Paths: vmhba0:C0:T0:L61, vmhba1:C0:T0:L61 

Notice the change in useANO which now reflects a value of 1.  In addition, I now have two Working Paths – an optimized path and an unoptimized path. 

I fired up ESXTOP and IOMETER which now showed a flurry of I/O traversing both paths.  I kid you not, it was a Clark Griswold moment when all the Christmas lights on the house finally worked.

So it took a while to figure this out but with some reading and the help of experts, I finally got it, and I was extremely jazzed.  What would have helped was if VMware’s PSA was more plug and play with various array types.  For instance, why can’t PSA recognize ALUA on the CLARiiON and automatically configure SATP for VMW_SATP_ALUA_CX?  Why is a reboot required for an SATP change?  PSA configuration in the vSphere client might have also been convenient but I recognize has diminishing returns or practical use with a large amount of hosts and/or LUNs to configure.  Scripting and CLI is the way to go for consistency and automation reasons or how about PSA configuration via Host Profiles? 

I felt a little betrayed and confused by the Navisphere GUI reflecting Failover Mode 1 after several attempts to change it to 4.  I was looking at host connectivity status. Was I looking in the wrong place? 

Lastly, end to end documentation on how to configure Round Robin would have helped a lot.  EMC got me part of the way there with the CLARiiON/VMware Applied Technology Guide document, but left me hanging, making no mention of the PSA configuration needed.  I’m getting that the end game for EMC multipathing today is PowerPath, which is fine – I’ll get to that, but I really wanted to do some testing with native Round Robin first, if for no other reason to establish a baseline to compare PowerPath to once I get there. 

Thanks again to the people I leaned on to help me through this.  It was the usual crew who can always be counted on.

VMTN Storage Performance Thread and the EMC Celerra NS-120

January 23rd, 2010

The VMTN Storage Performance Thread is a collaboration of storage performance results on VMware virtual infrastructure provided by VMTN Community members around the world.  The thread starts here, was locked due to length, and continues on in a new thread here.  There’s even a Google Spreadsheet version, however, activity in that data repository appears to have diminished long ago.  The spirit of the testing is outlined by thread creater and VMTN Virtuoso christianZ

“My idea is to create an open thread with uniform tests whereby the results will be all inofficial and w/o any warranty. If anybody shouldn’t be agreed with some results then he can make own tests and presents his/her results too. I hope this way to classify the different systems and give a “neutral” performance comparison. Additionally I will mention that the performance [and cost] is one of many aspects to choose the right system.” 

Testing standards are defined by christianZ so that results from each submission are consistent and comparable.  A pre-defined template is used in conjunction with IOMETER to generate the disk I/O and capture the performance metrics.  The test lab environment and the results are then appended to the thread discussion linked above.  The performance metrics measured are:

  1. Average Response Time (in Milliseconds, lower is better) – also known as latency of which VMware declares a potential problem threshold of 50ms in their Scalable Storage Performance whitepaper
  2. Average I/O per Second (number of I/Os, higher is better)
  3. Average MB per Second (in MB, higher is better)

Following are my results with the EMC Celerra NS-120 Unified Storage array

SERVER TYPE: Windows Server 2003 R2 VM ON ESXi 4.0 U1
CPU TYPE / NUMBER: VCPU / 1 / 1GB Ram (thin provisioned)
HOST TYPE: HP DL385 G2, 16GB RAM; 2x QC AMD Opteron 2356 Barcelona
STORAGE TYPE / DISK NUMBER / RAID LEVEL: EMC Celerra NS-120 / 15x 146GB 15K 4Gb FC / RAID 5
SAN TYPE / HBAs: Emulex dual port 4Gb Fiber Channel, HP StorageWorks 2Gb SAN switch
OTHER: Disk.SchedNumReqOutstanding and HBA queue depth set to 64 

Fibre Channel SAN Fabric Test

Test Name Avg. Response Time Avg. I/O per Second Avg. MB per Second
Max Throughput – 100% Read 1.62 35,261.29 1,101.92
Real Life – 60% Rand / 65% Read 16.71 2,805.43 21.92
Max Throughput – 50% Read 5.93 10,028.25 313.38
Random 8K – 70% Read 11.08 3,700.69 28.91
  
 
SERVER TYPE: Windows Server 2003 R2 VM ON ESXi 4.0 U1
CPU TYPE / NUMBER: VCPU / 1 / 1GB Ram (thin provisioned)
HOST TYPE: HP DL385 G2, 16GB RAM; 2x QC AMD Opteron 2356 Barcelona
STORAGE TYPE / DISK NUMBER / RAID LEVEL: EMC Celerra NS-120 / 15x 146GB 15K 4Gb FC / 3x RAID 5 5×146
SAN TYPE / HBAs: swISCSI
OTHER: Shared NetGear 1Gb SoHo Ethernet switch

swISCSI Test

Test Name Avg. Response Time Avg. I/O per Second Avg. MB per Second
Max Throughput – 100% Read 17.52 3,426.00 107.06
Real Life – 60% Rand / 65% Read 14.33 3,584.53 28.00
Max Throughput – 50% Read 11.33 5,236.50 163.64
Random 8K – 70% Read 15.25 3,335.68 22.06
  
 
SERVER TYPE: Windows Server 2003 R2 VM ON ESXi 4.0 U1
CPU TYPE / NUMBER: VCPU / 1 / 1GB Ram (thin provisioned)
HOST TYPE: HP DL385 G2, 16GB RAM; 2x QC AMD Opteron 2356 Barcelona
STORAGE TYPE / DISK NUMBER / RAID LEVEL: EMC Celerra NS-120 / 15x 146GB 15K 4Gb FC / 3x RAID 5 5×146
SAN TYPE / HBAs: NFS
OTHER: Shared NetGear 1Gb SoHo Ethernet switch

NFS Test

Test Name Avg. Response Time Avg. I/O per Second Avg. MB per Second
Max Throughput – 100% Read 17.18 3,494.48 109.20
Real Life – 60% Rand / 65% Read 121.85 480.81 3.76
Max Throughput – 50% Read 12.77 4,718.29 147.45
Random 8K – 70% Read 123.41 478.17 3.74

Please read further below for futher NFS testing results after applying EMC Celerra best practices

Fibre Channel Summary

Not surprisingly, Celerra over SAN fabric beats the pants off of the shared storage solutions I’ve had in the lab previously, HP MSA1000 and Openfiler 2.2 swISCSI before that, in all four IOMETER categories.  I was, however, pleasantly surprised to find that Celerra over fibre channel was one of the top performing configurations among a sea of HP EVA, Hitachi, NetApp, and EMC CX series frames.

swISCSI Summary

Celerra over swISCSIwas only slightly faster than the Openfiler 2.2 swISCSI on HP Proliant ML570 G2 hardware I had in the past on the Max Throughput-100%Read test. In the other three test categories, however, the Celerra left the Openfiler array in the dust.

NFS Summary

Moving on to Celerra over NFS, performance results were consistent with swISCSI in two test categories (Max Throughput-100%Read and Max Throughput-50%Read), but NFS performance numbers really dropped in the remaining two categories as compared to swISCSI (RealLife-60%Rand-65%Read and Random-8k-70%Read). 

What’s worth noting is that both the iSCSI and NFS datastores are backed by the same logical Disk Group and physical disks on the Celerra.  I did this purposely to compare the iSCSI and NFS protocols, with everything else being equal.  The differences in two out of the four categories are obvious.  The question came to mind:  Does the performance difference come from the Celerra, the VMkernel, or a combination of both?  Both iSCSI and NFS have evolved into viable protocols for production use in enterprise datacenters, therefore, I’m leaning AWAY from the theory that the performance degradation over NFS stems from the VMkernel. My initial conclusion here is that Celerra over NFS doesn’t perform as well with Random Read disk I/O patterns.  I welcome your comments and experience here.

Please read further below for futher NFS testing results after applying EMC Celerra best practices

CIFS

Although I did not test CIFS, I would like to take a look at its performance.  CIFS isn’t used directly by VMware virtual infrastructure, but it can be a handy protocol to leverage with NFS storage.  File management (ie. .ISOs, templates, etc.) on ESX NFS volumes becomes easier and more mobile and less tools are required when the NFS volumes are presented as CIFS shares on a predominantly Windows client network.  Providing adequate security through CIFS will be a must to protect the ESX datastore on NFS.

If you’re curious about storage array configuration and its impact on performance, cost, and availability, take a look at this RAID triangle which VMTN Master meistermn posted in one of the performance threads:

The Celerra stroage is currently carved out in the following way:

  0 1 2 3 4 5 6 7 8 9 10 11 12 13 14  
DAE 2 FC FC FC FC FC FC FC FC FC FC FC FC FC FC FC DAE 2
DAE1 NAS NAS NAS NAS NAS Spr Spr                 DAE 1
DAE 0 Vlt Vlt Vlt Vlt Vlt NAS NAS NAS NAS NAS NAS NAS NAS NAS NAS DAE 0
  0 1 2 3 4 5 6 7 8 9 10 11 12 13 14  

FC = fibre channel Disk Group

NAS = iSCSI/NFS Disk Groups

Spr = Hot Spare

Vlt = Celerra Valut drives

I’m very pleased with the Celerra NS-120.  With the first batch of tests complete, I’m starting to formulate ideas on when, where, and how to use the various storage protocols with the Celerra.  My goal is not to eliminate use of the slowest performing protocol in the lab.  I want to work with each of them on a continual basis to test future design and integration with VMware virtual infrastructure.

Update 1/30/10: New NFS performance numbers.  I’ve begun working with EMC vSpecialist to troubleshoot the performance descrepancies between swISCSI and NFS protocols.  A few key things have been identified and a new set of performance metrics have been posted below after making some changes:

  1. The first thing that the EMC vSpecialists (and others on the blog post comments) asked about was whether or not the file system uncached write mechanism was enabled. The uncached write mechanism is designed to improve performance for applications with many connections to a large file, such as a virtual disk file of a virtual machine.  This mechanism can enhance access to such large files through the NFS protocol.  Out of the box, the factory default is the uncached write mechanism is disabled on the Celerra. EMC recommends this feature be enabled with ESX(i).  The beauty here is that the feature can be toggled while the NFS file system is mounted on cluster hosts with VMs running on it.  VMware ESX Using EMC Celerra Storage Systems pages 99-101 outlines this recommendation.
  2. Per VMware ESX Using EMC Celerra Storage Systems pages 73-74, NFS send and receive buffers should be divisible by 32k on the ESX(i) hosts.  Again, these advanced settings can be adjusted on the hosts while VMs are running and the settings do not require a reboot.  EMC recommended a value of 64 (presumably for both).
  3. Use the maximum amount of write cache possible for Storage Processors (SPs). Factory defaults here:  598BM total read cache size, 32MB read cache size, 598MB total write cache size, 566MB write cache size.
  4. Specific to this test – verify that the ramp up time is 120 seconds.  Without the ramp up the results can be skewed. The tests I originall performed were with a 0 second ramp up time.

The new NFS performance tests are below, using some of the recommendations above: 

SERVER TYPE: Windows Server 2003 R2 VM ON ESXi 4.0 U1
CPU TYPE / NUMBER: VCPU / 1 / 1GB Ram (thin provisioned)
HOST TYPE: HP DL385 G2, 16GB RAM; 2x QC AMD Opteron 2356 Barcelona
STORAGE TYPE / DISK NUMBER / RAID LEVEL: EMC Celerra NS-120 / 15x 146GB 15K 4Gb FC / 3x RAID 5 5×146
SAN TYPE / HBAs: NFS
OTHER: Shared NetGear 1Gb SoHo Ethernet switch

New NFS Test After Enabling the NFS file system Uncached Write Mechanism

VMware ESX Using EMC Celerra Storage Systems pages 99-101

Test Name Avg. Response Time Avg. I/O per Second Avg. MB per Second
Max Throughput – 100% Read 17.39 3,452.30 107.88
Real Life – 60% Rand / 65% Read 20.28 2,816.13 22.00
Max Throughput – 50% Read 19.43 3,051.72 95.37
Random 8K – 70% Read 19.21 2,878.05 22.48
Significant improvement here!  
 
 
SERVER TYPE: Windows Server 2003 R2 VM ON ESXi 4.0 U1
CPU TYPE / NUMBER: VCPU / 1 / 1GB Ram (thin provisioned)
HOST TYPE: HP DL385 G2, 16GB RAM; 2x QC AMD Opteron 2356 Barcelona
STORAGE TYPE / DISK NUMBER / RAID LEVEL: EMC Celerra NS-120 / 15x 146GB 15K 4Gb FC / 3x RAID 5 5×146
SAN TYPE / HBAs: NFS
OTHER: Shared NetGear 1Gb SoHo Ethernet switch

New NFS Test After Configuring
NFS.SendBufferSize = 256 (this was set at the default of 264 which is not divisible by 32k)
NFS.ReceiveBufferSize = 128 (this was already at the default of 128)

VMware ESX Using EMC Celerra Storage Systems pages 73-74

Test Name Avg. Response Time Avg. I/O per Second Avg. MB per Second
Max Throughput – 100% Read 17.41 3,449.05 107.78
Real Life – 60% Rand / 65% Read 20.41 2,807.66 21.93
Max Throughput – 50% Read  18.25  3,247.21  101.48
Random 8K – 70% Read  18.55  2,996.54  23.41
Slight change  
 
 
SERVER TYPE: Windows Server 2003 R2 VM ON ESXi 4.0 U1
CPU TYPE / NUMBER: VCPU / 1 / 1GB Ram (thin provisioned)
HOST TYPE: HP DL385 G2, 16GB RAM; 2x QC AMD Opteron 2356 Barcelona
STORAGE TYPE / DISK NUMBER / RAID LEVEL: EMC Celerra NS-120 / 15x 146GB 15K 4Gb FC / 3x RAID 5 5×146
SAN TYPE / HBAs: NFS
OTHER: Shared NetGear 1Gb SoHo Ethernet switch

New NFS Test After Configuring IOMETER for 120 second Ramp Up Time

Test Name Avg. Response Time Avg. I/O per Second Avg. MB per Second
Max Throughput – 100% Read  17.28  3,472.43  108.51
Real Life – 60% Rand / 65% Read  21.05  2,726.38  21.30
Max Throughput – 50% Read  17.73  3,338.72  104.34
Random 8K – 70% Read  17.70  3,091.17  24.15

Slight change

Due to the commentary received on the 120 second ramp up, I re-ran the swISCSI test to see if that changeded things much.  To fairly compare protocol performance, the same parameters must be used across the board in the tests.

SERVER TYPE: Windows Server 2003 R2 VM ON ESXi 4.0 U1
CPU TYPE / NUMBER: VCPU / 1 / 1GB Ram (thin provisioned)
HOST TYPE: HP DL385 G2, 16GB RAM; 2x QC AMD Opteron 2356 Barcelona
STORAGE TYPE / DISK NUMBER / RAID LEVEL: EMC Celerra NS-120 / 15x 146GB 15K 4Gb FC / 3x RAID 5 5×146
SAN TYPE / HBAs: swISCSI
OTHER: Shared NetGear 1Gb SoHo Ethernet switch

New swISCSI Test After Configuring IOMETER for 120 second Ramp Up Time

Test Name Avg. Response Time Avg. I/O per Second Avg. MB per Second
Max Throughput – 100% Read  17.79  3,351.07  104.72
Real Life – 60% Rand / 65% Read  14.74  3,481.25  27.20
Max Throughput – 50% Read  12.17  4,707.39  147.11
Random 8K – 70% Read  15.02  3,403.39  26.59

swISCSI is still performing slightly better than NFS on the Random Reads, however, the margin is much closer

At this point I am content, stroke, happy, (borrowing UK terminology there) with NFS performance.  I am now moving on to ALUA, Round Robin, and PowerPath/VE testing.  I set up NPIV over the weekend with the Celerra as well – look for a blog post coming up on that.

Thank you EMC and to the folks who replied in the comments below with your help tackling best practices and NFS optimization/tuning!