Symptoms
ESXi 5.x => ESXi 6.x
How to properly tune an ESX server that is connected to a Symmetrix.
Degraded performance is seen on new or existing ESX servers zoned to Symmetrix storage arrays.
Degraded performance on an ESX server attached to a DMX, VMAX, or VMAX3 array.
Note that this does not solely apply to newly implemented servers.
Cause
A tunable setting for VMware' s Native Multipathing (NMP) Round Robin (RR) policy is incorrectly set or is still set to the default value of 1000.
Resolution
The NMP RR policy I/O operations limit should be set to a value of "1" when attached to a Symmetrix.
For ESX/ESXi 4.x:
- To check the I/O operations limit: esxcli nmp roundrobin getconfig --device=<device NAA>
- To set the I/O operations limit: esxcli nmp roundrobin setconfig --device=<device NAA> --iops 1 --type iops
For ESXi 5.x:
- To check the I/O operation limit: esxcli storage nmp psp roundrobin deviceconfig get --device=<device NAA>
- To set the I/O operations limit: esxcli storage nmp psp roundrobin deviceconfig set --device=<device NAA> --iops=1 --type iops
Prior to VMware ESXi 5.1 the default path selection policy (PSP) for the Symmetrix array type plug-in (SATP) was fixed. The PSP for Symmetrix SATP should be changed to a Round Robin policy (the default policy in ESXi 5.1). When this change is applied, NMP channels I/O in a Round Robin fashion across all active paths.
However, the frequency of the path switching is controlled by a parameter called "i/o operation limit" and by default is set to 1000. This is incorrect when zoned to an EMC Symmetrix only.
Refer to EMC Knowledgebase article 13243
How to change the NMP policy in VMware ESX Server 4 to Round Robin.
Before engaging the performance teams at EMC, the following article should be reviewed and applied. This tunable affects I/O performance on all ESX servers attached to Symmetrix storage arrays.
Affected Products
VMware ESX
Products
VMware