Ubiquitous Talk

Running ZFS over NFS as a VMware Store

NFS is definitely a very well rounded high performance file storage system and it certainly serves VMware Stores successfully over many storage products. Recently one of my subscribers asked me if there was a reason why my blogs were more centric to iSCSI. Thus the question was probing for a answer to a question many of us ask ourselves. Is NFS superior to block based iSCSI and which one should I choose for VMware. The answer to this question is not which protocol is superior but which protocol serves to provision the features and function you require most effectively. I use both protocols and find they both have desirable capability and functionality and conversely have some negative points as well.

NFS typically is generally more accessible because its a file level protocol and sits higher up on the network stack. This makes it very appealing when working with VMware virtual disks aka vmdk’s simply because they also exist at the same layer. NFS is ubiquitous across NAS vendors and can be provisioned by multiple agnostic implementation endpoints.  An NFS protocol hosts the capability to be virtualized and encapsulated within any Hypevisor instance either clustered or standalone. The network file locking and share semantics of NFS grant it a multitude of configurable elements which can serve a wide range of applications.

In this blog entry we will explore how to implement an NFS share for VMware ESX using OpenSolaris and ZFS. We will also explore a new way of accelerating the servers I/O performance with a new product called the DDRdrive X1.

OpenSolaris is an excellent choice for provisioning NFS storage volumes on VMware.  It hosts many advanced desirable storage features that set it far ahead of other Unix flavors. We can use the advanced networking features and ZFS including the newly integrated dedup functionality to craft the best NFS functionality available today.

Let start by examining the overall NAS storage architecture.

In this architecture we are defining a fault tolerant configuration using two physical 1Gbe switches with a quad or dual Ethernet adapter(s). On the OpenSolaris storage head we are using IPMP aka IP Multipathing to establish a single IP address to serve our NFS store endpoint. A single IP is more appropriate for VMware environments as they do not support multiple NFS IP targets per NFS mount point.  IPMP provisions layer 3 load balancing and interface fault tolerance. IPMP commonly uses ICMP and default routes to determine interface failure states thus it well suited for a NAS protocol service layer. In a effort to reduce excessive ICMP rates we will aggregate the two dual interfaces into a single channel connection to each switch. This will allow us to define two test IP addresses for the IPMP service and keep our logical interface count down to a minimum. We are also defining a 2 port trunk/aggregate between the two physical switches which provides more path availability and reduces  switch failure detection times.

On the ESX host side we are defining 1 interface per switch. This type of configuration requires that only one of the VMware interfaces is an active team member vmnic within a single vSwitch definition. If this is not configured this way the ESX host will fail to detect and activate the second nic under some failure modes. This is not a bandwidth constraint issue since the vmkernel IP interface will only activity use one nic.

With an architecture set in place let now explore some of the pros and cons of running VMware on Opensolaris NFS.

Some of the obvious pros are:

• VMware uses NFS in a thin provisioned format.
• VMDKs are stored as files and are mountable over a variety of hosts.
• Simple backup and recovery.
• Simple cloning and migration.
• Scalable storage volumes.

And some of the less obvious pros:

• IP based transports can be virtualized and encapsulated for disaster recovery.
• No vendor lock-in
• ZFS retains NFS share properties within the ZFS filesystem.
• ZFS will dedup VMDKs files at the block level.

And there are the cons:

• Every write I/O from VMware is an O_SYNC write.
• Firewall setups are complex.
• Limited in its application. Only NFS clients can consume NFS file systems.
• General  protocol security challenges. (RPC)
• VMware kernel constraints
• High CPU overhead.
• Bursty data flow.

Before we break out into the configuration detail level lets examine some of the VMware and NFS behaviors so as to gain some insight into the reason I primarily use iSCSI for most VMware implementations.

I would like demonstrate some characteristics that are primarily a VMware client side behavior and it’s important that you are aware of them when your considering NFS as a Datastore.

This VMware performance chart of an IOMeter generated load reveals the burst nature of the NFS protocol. The VMware NFS client exclusively uses a O_SYNC flag on write operations which requires a committed response for the NFS server. At some point the storage system will not be able to complete every request and thus a pause in transmission will occur. The same occurs on reads when the network component buffers reach saturation. In this example chart we are observing a single 1Gbe interface at saturation from a read stream.

In this output we are observing a read stream across vh0 which is one of two active ESX4 host VMs loading our OpenSolaris NFS store and we can see the maximum network throughput is achieved which is ~81MB/s. If you examine the average value of 78MB/s you can see the burst events do not have significant impact and is not a bandwidth concern with ~3MB/s of loss.

At the same time we are recording this write stream chart on vh3 a second ESX 4 host loading the same NFS OpenSolaris store.  As I would expect, its very similar to the read stream except that we can see the write performance is lower and that’s to be expected with any write operations. We can also identify that we are using a full duplex path transmission across to our OpenSolaris NFS host since vh0 is reading (recieving) and vh3 is writing(transmitting).

In this chart we are observing a limiting characteristic of the VMware vmkernel NFS client process. We have introduced a read stream in combination with a preexisting active write stream on a single ESX host. As you can see the transmit and receive packet rates are both reduced and now sum to a maximum of ~75MB/s.

Transitioning from read to write active streams confirms the transmission is limited to ~75Mb/s regardless the full duplex interface capability.  This information demonstrates that a host using 1Gbe ethernet connections will be constrained based on its available resources. This is a important element to consider when using NFS as a VMware datastore.

Another important element to consider is the CPU load impact of running the vmkernel NFS client. There is a significant CPU cycle cost on VMware hosts and this is very apparent under heavier loads. The following screen shot depicts a running IOmeter load test against our OpenSolaris NFS store. The important elements are as follows. IOMeter is performing 32KB reads in a 100% sequential access mode which drives a CPU load on the VM of ~35% however this is not the only CPU activity that occurs for this VM.

When we examine the ESX host resource summary for the running VM we can now observe the resulting overhead load which is realized by viewing the Consumed Host CPU value. The VM in this case is granted 2 CPUs each are a 3.2Ghz Intel hypervisor resource. We can see that the ESX host is running at 6.6Ghz to drive the vmkernel NFS I/O load.

Lets see the performance chart results when we svMotion the activily loaded running VM on the same ESX host to an iSCSI VMFS based store on the same OpenSolaris storage host. The only elements changing in this test are the underlying storage protocols. Here we can clearly see CPU object 0 is the ESX host CPU load. During the svMotion activity we begin to see some I/O drop off due to the addition background disk load. Finally we observe the VM transition at the idle point and the resultant CPU load of iSCSI I/O impact. We clearly see the ESX host CPU load drop from 6.6Ghz to 3.5Ghz which makes it very apparent the NFS requires substantially higher CPU that iSCSI.

With the svMotion completed we now observe the same IOMeter screen shot retake and its very obvious that our throughput and IOPS have increased significantly and the VM granted CPU load has not changed significantly.   A decrease of ESX host CPU load in the order of  ~55% and and increase of ~32% in IOPS and 45% of throughput shows us there are some negative behaviors to be cognizant of. Keep in mind that this is not that case when the I/O type is small and random like that of a Database in those cases  NFS is normally the winner, however VMware normally hosts mixed loads and thus we need to consider this negative effect at design time and when targeting VM I/O characteristics.

With a clear understanding of some important negative aspects to implementing NFS for VMware ESX hosts we can proceed to the storage system build detail. The first order of business is the hardware configuration detail. This build is simply one of my generic white boxes and it hosts the following hardware:

GA-EP45-DS3L Mobo with an Intel 3.2Ghz E8500 Core Duo

1 x 70GB OS Disk

2 x 500GB SATA II ST3500320AS disks

2GB of Ram

1 x Intel Pro 1000 PT Quad Network Adapter

As a very special treat on this configuration I am also privileged to run an DDRDrive X1 Cache Accelerator which I am currently testing some newly developed beta drivers for OpenSoalris. Normally I would use 4GB of ram as a minimum but I needed to constraint this build in a effort to load down the dedicated X1 LOG drive and the physical SATA disks thus this instance is running only 2GB of ram. In this blog entry I will not be detailing the OpenSolaris install process, we will begin from a Live CD installed OS.

OpenSolaris will default to a dynamic network service configuration named nwam, this needs to be disabled and the physical:default service enabled.

root@uss1:~# svcadm disable svc:/network/physical:nwam
root@uss1:~# svcadm enable svc:/network/physical:default

To establish an aggregation we need to un-configure any interfaces that we previously configured before proceeding.

root@uss1:~# ifconfig -a
lo0: flags=2001000849<UP,LOOPBACK,RUNNING,MULTICAST,IPv4,VIRTUAL> mtu 8232 index 1
inet 127.0.0.1 netmask ff000000
e1000g0: flags=1000843<UP,BROADCAST,RUNNING,MULTICAST,IPv4> mtu 1500 index 2
inet 10.1.0.1 netmask ffff0000 broadcast 10.255.255.255
ether 0:50:56:bf:11:c3
lo0: flags=2002000849<UP,LOOPBACK,RUNNING,MULTICAST,IPv6,VIRTUAL> mtu 8252 index 1
inet6 ::1/128

root@uss1:~# ifconfig e1000g0 unplumb

Once cleared the assignment of the physical devices is possible using the following commands

dladm create-aggr –d e1000g0 –d e1000g1 –P L2,L3 1
dladm create-aggr –d e1000g2 –d e1000g3 –P L2,L3 2

Here we have set the policy allowing layer 2 and 3 and defined two aggregates aggr1 and aggr2. We can now define the VLAN based interface shown here as VLAN 500 instances 1 are 2 respective of the aggr instances. You just need to apply the following formula for defining the VLAN interface.

(Adaptor Name) + vlan * 1000 + (Adaptor Instance)

ifconfig aggr500001 plumb up 10.1.0.1 netmask 255.0.0.0
ifconfig aggr500002 plumb up 10.1.0.2 netmask 255.0.0.0

Each pair of interfaces needs to be attached to a trunk definition on its switch path. Typically this will be a Cisco or HP switch in most environments. Here is a sample of how to configure each brand.

Cisco:

configure terminal
interface port-channel 1
interface ethernet 1/1
channel-group 1
interface ethernet 1/2
channel-group 1
interface ethernet po1
switchport mode trunk allowed vlan 500
exit

HP Procurve:

trunk 1-2 trk1 trunk
vlan 500
name “eSAN1”
tagged trk1

 

Once we have our two physical aggregates setup we can define the IP multipathing interface components.  As a best practice we should define the IP addresses in our hosts file and then refer to those names in the remaining configuration tasks.

Edit /etc/hosts to have the following host entries.

::1 localhost
127.0.0.1 uss1.local localhost loghost
10.0.0.1 uss1 uss1.domain.name
10.1.0.1 uss1.esan.data1
10.1.0.2 uss1.esan.ipmpt1
10.1.0.3 uss1.esan.ipmpt2

Here we have named the IPMP data interface aka a public IP as uss1.esan-data1 this ip will be the active connection for our VMware storage consumers.  The other two named uss1.esan-ipmpt1 and uss1.esan-ipmpt2 are beacon probe  IP test addresses and will not be available to external connections.

IPMP functionallity is included with OpenSolaris and is configured with the ifconfig utility. The follow sets up the first aggregate with a real public IP and a test address. The deprecated keyword defines the IP as a test address and the failover keyword defines if the IP can be moved in the event of interface failure.

ifconfig aggr500001 plumb uss1.esan.ipmpt1 netmask + broadcast + group ipmpg1 deprecated -failover up addif uss1.esan.data1 netmask + broadcast + failover up
ifconfig aggr500002 plumb uss1.esan.ipmpt2 netmask + broadcast + group ipmpg1 deprecated -failover up

To persist the IPMP network configuration on boot you will need to create hostname files matching the interface names with the IPMP configuration statement store in them. The following will address it.

echo uss1.esan.ipmpt1 netmask + broadcast + group ipmpg1 deprecated -failover up addif uss1.esan.data1 netmask + broadcast + failover up > /etc/hostname.aggr500001

echo uss1.esan.ipmpt1 netmask + broadcast + group ipmpg1 deprecated -failover up > /etc/hostname.aggr500002

The resulting interfaces will look like the following:

root@uss1:~# ifconfig -a
lo0: flags=2001000849<UP,LOOPBACK,RUNNING,MULTICAST,IPv4,VIRTUAL> mtu 8232 index 1
inet 127.0.0.1 netmask ff000000
aggr1: flags=9040843<UP,BROADCAST,RUNNING,MULTICAST,DEPRECATED,IPv4,NOFAILOVER> mtu 1500 index 2
inet 10.1.0.2 netmask ff000000 broadcast 10.255.255.255
groupname ipmpg1
ether 0:50:56:bf:11:c3
aggr2: flags=9040843<UP,BROADCAST,RUNNING,MULTICAST,DEPRECATED,IPv4,NOFAILOVER> mtu 1500 index 3
inet 10.1.0.3 netmask ff000000 broadcast 10.255.255.255
groupname ipmpg1
ether 0:50:56:bf:6e:2f
ipmp0: flags=8001000843<UP,BROADCAST,RUNNING,MULTICAST,IPv4,IPMP> mtu 1500 index 5
inet 10.1.0.1 netmask ff000000 broadcast 10.255.255.255
groupname ipmpg1
lo0: flags=2002000849<UP,LOOPBACK,RUNNING,MULTICAST,IPv6,VIRTUAL> mtu 8252 index 1
inet6 ::1/128

In order for IPMP to detect failures in this configuration you will need to define target probe addresses for IPMP use. For example I use multiple ESX hosts as probe target on the storage network.

e.g.

root@uss1:~# route add -host 10.1.2.1 10.1.2.1 -static
root@uss1:~# route add -host 10.1.2.2 10.1.2.2 -static

This network configuration yields 2,2Gbe aggregate paths bound to a single logical active IP address on 10.1.0.1, with  interfaces aggr1 and aggr2 the keyword deprecated directs the IPMP mpathd service daemon to prevent application session connection packets establishment and the nofailover keyword instructs mpathd not to allow the bound IP to failover to any other interface in the IPMP group.

There are many other possible configurations but I prefer this method because it remains logically easy to diagnose and does not introduce unnecessary complexity.

Now that we have layer 3 network connectivity we should establish the other essential OpenSolaris static TCP/IP configuration elements. We need to ensure we have a persistent default gateway and our DNS client resolution enabled.

The persistent default gateway is very simple to define as is done with the route utility command as follows.

root@uss1:~# route -p add default 10.1.0.254
add persistent net default: gateway

When using NFS I prefer provisioning name resolution as a additional layer of access control. If we use names to define NFS shares and clients we can externally validate the incoming IP  with a static file or DNS based name lookup. An OpenSolaris NFS implementation inherently grants this methodology.  When a client IP requests access to an NFS share we can define a forward lookup to ensure the IP maps to a name which is granted access to the targeted share. We can simply define the desired FQDNs against the NFS shares.

In small configurations static files are acceptable as is in the case here. For large host farms the use of a DNS service instance would ease the admin cycle. You would just have to be careful that your cached TimeToLive (TTL) value is greater that 2 hours thus preventing excessive name resolution traffic. The TTL value will control how long the name is cached and this prevents constant external DNS lookups.

To configure name resolution for both file and DNS we simply copy the predefined config file named nsswitch.dns to the active config file nsswitch.conf as follows:

root@uss1:~# cp /etc/nsswitch.dns /etc/nsswitch.conf

Enabling DNS will require the configuration of our /etc/resolv.conf file which defines our name servers and namespace.

e.g.

root@ss1:~# cat /etc/resolv.conf
domain laspina.ca
nameserver 10.1.0.200
nameserver 10.1.0.201

You can also use the static /etc/hosts file to define any resolvable name to IP mapping.

With OpenSolaris you should always define your NFS share properties using the ZFS administrative tools. When this method is used we can the take advantage of keeping the NFS share properties inside of ZFS. This is really useful when you replicate or clone the ZFS file system to an alternate host as all the share properties will be retained. Here are the basic elements of an NFS share configuration for use on VMware storage consumers.

zfs create -p sp1/nas/vol1
zfs set mountpoint=/export/uss1-nas-vol1 sp1/nas/vol1
zfs set sharenfs=rw,nosuid,root=vh3-nas:vh2-nas:vh1-nas:vh0-nas sp1/nas/vol1

The ACL NFS share property of rw sets the entire share as read write, you could alternately use rw=hostname for each host but it seems redundant to me.  The nosuid prevents any incoming connection from switching user ids for example from a non-root value to 0. Finally the root=hostname property grants the incoming host name access to the share with root access permissions.  Any files created by the host will be as the root id. While these steps are some level of access control it falls well short of secure thus I also keep the NAS subnets fully isolated or firewalled to prevent external network access to the NFS share hosts.

Once our NFS share is up and running we can proceed to configure the VMware network components and share connection properties. VMware requires a vmkernel network interface definition to provision NFS connectivity. You should dedicate a vmnic team and a vswitch for your storage network.

Here is a visual  example of a vmkernel configuration with a teamed pair of vmnics

As you can see we have dedicated the vSwitch and vmnics on VLAN 500, no other traffic should be permitted on this network. You should also set the default vmkernel gateway to its own address. This will promote better performance as there is no need to leave this network.

For eNAS-Interface1 you should define one active and one standby vmnic. This will ensure proper interface fail-over in all failure modes.  The VMware NFS kernel instance will only use a single vmnic so your not loosing any bandwidth. The vmnic team only serves as a fault tolerant connection and is not a load balanced configuration.

At this point you should validate your network connectivity by pinging the vmkernel IP address from the OpenSolaris host. If you chose to ping from ESX use vmkping instead of ping otherwise you will not get a response.

Provided your network connectivity is good you can define your vmkernel NFS share properties. Here is a visual example.

And if you prefer an ESX command line method:

esxcfg-nas -a -o uss1-nas -s /export/uss1-nas-vol1 uss1-nas-vol1

In this example we are using a DNS based name of uss1-nas. This would allow you to change the host IP without having to reconfigure VMware hosts. You will want to make sure the DNS name cache TTL in not a small value for two reasons. One an DNS outage would impact the IP resolution and as well you do not want excessive resolution traffic on the eSAN subnet(s).

The NFS share configuration info is maintained in the /etc/vmware/esx.conf file and looks like the following example.

/nas/uss1-nas-vol1/enabled = “true”
/nas/uss1-nas-vol1/host = “uss1-nas”
/nas/uss1-nas-vol1/readOnly = “false”
/nas/uss1-nas-vol1/share = “/export/uss1-nas-vol1”

If your trying to change NFS share parameters and the NFS share is not available after a successful configuration you could run into a messed up vmkernel NFS state and you’ll receive the following message:

Unable to get Console path for Mount

You will need to reboot the ESX server to clean it up so don’t mess with anything else until that is performed. (I’ve wasted a few hours on that buggy VMware kernel NFS client behavior).

Once the preceeding steps are successful the result will be a NAS based NFS share which is now available like this example.

With a working NFS storage system we can now look at optimizing the I/O capability of ZFS and NFS.

VMware performs write operations over NFS using an O_SYNC control flag. This will force the storage system to commit all write operations to disk to ensure VM file integrity. This can be very expensive when it comes to high performance IOPS especially when using SATA architecture. We could disable our ZIL aka ZFS Intent Log but this could result in severe corruption in the event of a systems fault or environmental issue. A much better alternative is to use a non-volatile ZIL device. In this case we have an DDRdrive X1 which is a 4GB high speed externally powered dram bank with a high speed SCSI interface and also hosts 4GB of flash for long term shutdowns.  The DDRdrive X1 IO capability reaches the 200,000/sec range and up. By using an external UPS power source we can economically prevent ZFS corruption and reap the high speed benefits of dram even when unexpected system interruptions occur.

In this blog our storage host is using Seagate ST3500320AS disk which are challenged to achieve ~180 IOPS. And that IO rate is under ideal sequential read write loads. With a cache we can expect that these disks will deliver no greater than 360 IOPS under ideal conditions.

Now lets see if this is true based on some load tests using Microsoft’s SQLIO tool. First we will disable our ZFS ZIL caching DDRdrive X1 show here as device c9t0d0

NAME        STATE     READ WRITE CKSUM
sp1         DEGRADED     0     0     0
mirror-0  ONLINE       0     0     0
c6t1d0  ONLINE       0     0     0
c6t2d0  ONLINE       0     0     0
logs
c9t0d0  OFFLINE      0     0     0

No lets run the SQLIO test for 5 minutes with random 8K I/O write requests which are simply brutal for any SATA disk to keep up with.  We have defined a file size of 32GB to ensure we hit the disk by exceeding our 2GB cache memory foot print. As you can see from the output we achieve 227 IOs/sec which is below the mirrored drive pair capability.

C:Program FilesSQLIO>sqlio -kW -s300 -frandom -o4 -b8 -LS -Fparam.txt
sqlio v1.5.SG
using system counter for latency timings, 3579545 counts per second
parameter file used: param.txt
file c:testfile.dat with 2 threads (0-1) using mask 0x0 (0)
2 threads writing for 300 secs to file c:testfile.dat
using 8KB random IOs
enabling multiple I/Os per thread with 4 outstanding
using specified size: 32768MB for file: c:testfile.dat
initialization done
CUMULATIVE DATA:
throughput metrics:
IOs/sec:   227.76
MBs/sec:     1.77
latency metrics:
Min_Latency(ms): 8
Avg_Latency(ms): 34
Max_Latency(ms): 1753
histogram:
ms: 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24+
%:  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1  1 29  7  3  2  1  1  1 54

new  name   name  attr  attr lookup rddir  read read  write write
file remov  chng   get   set    ops   ops   ops bytes   ops bytes
0     0     0   300     0      0     0     3   16K   146 1.12M /export/uss1-nas-vol1
0     0     0   617     0      0     0     0     0   309 2.39M /export/uss1-nas-vol1
0     0     0   660     0      0     0     0     0   329 2.52M /export/uss1-nas-vol1
0     0     0   677     0      0     0     0     0   338 2.63M /export/uss1-nas-vol1
0     0     0   638     0      0     0     0     0   321 2.46M /export/uss1-nas-vol1
0     0     0   496     0      0     0     0     0   246 1.88M /export/uss1-nas-vol1
0     0     0    44     0      0     0     0     0    21  168K /export/uss1-nas-vol1
0     0     0   344     0      0     0     0     0   172 1.32M /export/uss1-nas-vol1
0     0     0   646     0      0     0     0     0   323 2.51M /export/uss1-nas-vol1
0     0     0   570     0      0     0     0     0   285 2.20M /export/uss1-nas-vol1
0     0     0   695     0      0     0     0     0   350 2.72M /export/uss1-nas-vol1
0     0     0   624     0      0     0     0     0   309 2.38M /export/uss1-nas-vol1
0     0     0   562     0      0     0     0     0   282 2.15M /export/uss1-nas-vol1

Now lets enable the DDRdrive X1 ZIL cache and see where that takes us.

NAME        STATE     READ WRITE CKSUM
sp1         ONLINE       0     0     0
mirror-0  ONLINE       0     0     0
c6t1d0  ONLINE       0     0     0
c6t2d0  ONLINE       0     0     0
logs
c9t0d0  ONLINE       0     0     0

Again we run the identical SQLIO test and results are dramatically different, we immediately see a 4X improvement in IOPS but whats much more important is the reduction in latency which will make any database workload fly.

C:Program FilesSQLIO>sqlio -kW -s300 -frandom -o4 -b8 -LS -Fparam.txt
sqlio v1.5.SG
using system counter for latency timings, 3579545 counts per second
parameter file used: param.txt
file c:testfile.dat with 2 threads (0-1) using mask 0x0 (0)
2 threads writing for 300 secs to file c:testfile.dat
using 8KB random IOs
enabling multiple I/Os per thread with 4 outstanding
using specified size: 32768 MB for file: c:testfile.dat
initialization done
CUMULATIVE DATA:
throughput metrics:
IOs/sec:   865.75
MBs/sec:     6.76
latency metrics:
Min_Latency(ms): 0
Avg_Latency(ms): 8
Max_Latency(ms): 535
histogram:
ms: 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24+
%: 56 13  9  3  1  0  1  1  1  1  1  1  1  1  1  0  0  0  0  0  0  0  0  0  7

new  name   name  attr  attr lookup rddir  read read  write write
file remov  chng   get   set    ops   ops   ops bytes   ops bytes
0     0     0   131     0      0     0     0     0    66  516K /export/uss1-nas-vol1
0     0     0 3.23K     0      0     0     0     0 1.62K 12.8M /export/uss1-nas-vol1
0     0     0    95     0      0     0     2    8K    43  324K /export/uss1-nas-vol1
0     0     0 2.62K     0      0     0     0     0 1.31K 10.3M /export/uss1-nas-vol1
0     0     0   741     0      0     0     0     0   369 2.78M /export/uss1-nas-vol1
0     0     0 1.99K     0      0     0     0     0  1019 7.90M /export/uss1-nas-vol1
0     0     0 1.34K     0      0     0     0     0   687 5.32M /export/uss1-nas-vol1
0     0     0   937     0      0     0     0     0   468 3.62M /export/uss1-nas-vol1
0     0     0 2.60K     0      0     0     0     0 1.30K 10.3M /export/uss1-nas-vol1
0     0     0 2.02K     0      0     0     0     0 1.01K 7.84M /export/uss1-nas-vol1
0     0     0 1.91K     0      0     0     0     0   978 7.58M /export/uss1-nas-vol1
0     0     0 1.94K     0      0     0     0     0   992 7.67M /export/uss1-nas-vol1

When we look at ZFS ZIL caching devices there are some important elements to consider. For most provisioned VMware storage systems you do not require large volumes of ZIL cache to generate good I/O performance.  What you need to do is carefully determine the active data write footprint size. Remember that ZIL is a write only world and that those writes will be relocated to a slower disk at some point. These relocation functions are processed in batches or as Ben Rockwood likes to say in a regular breathing cycle.  This means that random I/O operations can queued up and converted to a more sequential like behavior characteristic. Random synchronous write operations can be safely acknowledged immediately and then the ZFS DMU can process them more efficiently in the background. This means that if we provision cache devices that are closer to the system bus and have lower latency the back end core compute hardware will be able to move the data ahead of the bursting I/O peak up ramps and thus we deliver higher IOPS with significantly less cache requirements. Devices like the DDRdrive X1 are a good example of implementing this strategy.

I hope you found this blog entry to be interesting and useful.

Regards,

Mike

Site Contents: © 2010  Mike La Spina

Running ZFS over iSCSI as a VMware vmfs store

The first time I looked at ZFS it totally floored me. This is a file system that has changed storage system rules as we currently know them and continues to do so. It is with no doubt the best architecture to date and now you can use it for your VMware stores.

Previously I had explored using it for a VMware store but ran into many issues which were real show stoppers. Like the VPD page response issue which made VMware see only one usable iSCSI store. But things are soon to be very different when Sun releases the snv_93 or above to all. I am currently using the unreleased snv_93 iscsitgt code and it works with VMware in all the ways you would want. Many thanks to the Sun engineers for adding NAA support on the iSCSI target service. With that being said let me divulge the details and behaviors of the first successful X4500 ZFS iSCSI VMware implementation in the real world.

Lets look at the Architectural view first.

 
The architecture uses a best practice approach consisting of completely separated physical networks for the iSCSI storage data plane. All components have redundant power and network connectivity. The iSCSI storage backplane is configured with an aggregate and is VLAN’d off from the server management network. Within the physical HP 2900’s an inter-switch ISL connection is defined but is not critical. This allows for more available data paths if additional interfaces were assigned on the ESX host side.
The Opensolaris aggregate and network components are configured as follows:

For those of you using Indiana….By default nwam is enabled on Indiana and this needs to be disabled and the physical network service enabled.

svcadm disable svc:/network/physical:nwam
svcadm enable svc:/network/physical:default

The aggregate is defined using the data link adm utility but first any bindings need to be cleared by unplumbing the interfaces.

e.g. ifconfig e1000g0 unplumb

Once cleared the assignment of the physical devices is possible using the following commands

dladm create-aggr –d e1000g0 –d e1000g1 –P L2,L3 1
dladm create-aggr –d e1000g2 –d e1000g3 –P L2,L3 2

Here we have set the policy allowing layer 2 and 3 and defined two aggregates aggr1 and aggr2. We can now define the VLAN based interface shown here as VLAN 500 instances 1 are 2 respective of the aggr instances. You just need to apply the following formula for defining the VLAN interface.

(Adaptor Name) + vlan * 1000 + (Adaptor Instance)

ifconfig aggr500001 plumb up 10.1.0.1 netmask 255.255.0.0
ifconfig aggr500002 plumb up 10.1.0.2 netmask 255.255.0.0

To persist the network configuration on boot you will need to create hostname files and hosts entries for the services to apply on startup.

echo ss1.iscsi1 > /etc/hostname.aggr500001
echo ss1.iscsi2 > /etc/hostname.aggr500002

Edit /etc/hosts to have the following host entries.

::1 localhost
127.0.0.1 ss1.local localhost loghost
10.0.0.1 ss1 ss1.domain.name
10.1.0.1 ss1.iscsi1
10.1.0.2 ss1.iscsi2

On the HP switches its a simple static trunk definition on port 1 and 2 using the following at the CLI.

trunk 1-2 trk1 trunk 

Once all the networking components are up and running and persistent, its time to define the ZFS store and iSCSI targets. I chose to include both mirrored and raidz pools. I needed to find and organize the cxtxdx device names using the cfgadm command or you could issue a format command as well to see the controller, target, disk names if you’re not using an X4500. I placed the raidz devices across controllers to improve I/O and distribute the load. It would not be a prudent to place one array on a single SATA controller. So here is what it ends up looking like from the ZFS command view.

zpool create –f rp1 raidz1 c4t0d0 c4t6d0 c5t4d0 c8t2d0 c9t1d0 c10t1d0
zpool add rp1 raidz1 c4t1d0 c4t7d0 c5t5d0 c8t3d0 c9t2d0 c10t2d0
zpool add rp1 raidz1 c4t2d0 c5t0d0 c5t6d0 c8t4d0 c9t3d0 c10t3d0
zpool add rp1 raidz1 c4t3d0 c5t1d0 c5t7d0 c8t5d0 c9t5d0 c11t0d0
zpool add rp1 raidz1 c4t4d0 c5t2d0 c8t0d0 c8t6d0 c9t6d0 c11t1d0
zpool add rp1 raidz1 c4t5d0 c5t3d0 c8t1d0 c8t7d0 c10t0d0 c11t2d0
zpool add rp1 spare c11t3d0
zpool create –f mp1 mirror c10t4d0 c11t4d0
zpool add mp1 mirror c10t5d0 c11t5d0
zpool add mp1 mirror c10t6d0 c11t6d0
zpool add mp1 spare c9t7d0

It only takes seconds to create terabytes of storage, wow it truly is a thing of beauty (geek!). Now it’s time to define a few pools and stores in preparation for the creation of the iSCSI targets. I chose to create units of 750G since VMware would not perform well with much more than that. This is somewhat dependant on the size of the VM and type of I/O but generally ESX host will serve a wide mix so try I keep it to a reasonable size or it ends up with SCSI reservation issues (that’s a bad thing chief).

You must also consider I/O block size before creating a ZFS store this is not something that can be changed later so now is the time. It’s done by adding the –b 64K to the ZFS create command. I chose to use 64k for the block size which aligns with VMWare default allocation size thus optimizing performance. The –s option enables a sparse volume feature aka thin provisioning. In this case the space was available but it is my favorite way to allocate storage.

zfs create rp1/iscsi
zfs create -s -b 64K -V 750G rp1/iscsi/lun0
zfs create -s -b 64K -V 750G rp1/iscsi/lun1
zfs create -s -b 64K -V 750G rp1/iscsi/lun2
zfs create -s -b 64K -V 750G rp1/iscsi/lun3
zfs create mp1/iscsi
zfs create -s -b 64K -V 750G mp1/iscsi/lun0

Originally I wanted to build the ESX hosts using a local disk but thanks to some bad IBM x346 engineering I could not use the QLA4050C and an integrated Adaptec controller on the ESX host server hardware. So I decided to give boot from iSCSI a go thus here is the boot LUN definition that I used for it. The original architectural design requires local disk to prevent an ESX host failure in the event of an iSCSI path outtage.

zfs create rp1/iscsi/boot
zfs create -s -V 16G rp1/iscsi/boot/esx1

Now that the ZFS stores are complete we can create the iSCSI targets for the ESX hosts to use. I have named the target alias to reflect something about the storage system which makes it easier to work with. I also created an iSCSI configuration store so we can persist the iSCSI targets on reboots. (This may now be included with Opensolaris Indiana but I have not tested it)

mkdir /etc/iscsi/config
iscsitadm modify admin –base-directory /etc/iscsi/config
iscsitadm create target -u 0 -b /dev/zvol/rdsk/rp1/iscsi/lun0 ss1-zrp1
iscsitadm create target -u 1 -b /dev/zvol/rdsk/rp1/iscsi/lun1 ss1-zrp1
iscsitadm create target -u 2 -b /dev/zvol/rdsk/rp1/iscsi/lun2 ss1-zrp1
iscsitadm create target -u 3 -b /dev/zvol/rdsk/rp1/iscsi/lun3 ss1-zrp1
iscsitadm create target -b /dev/zvol/rdsk/mp1/iscsi/lun0 ss1-zmp1
iscsitadm create target -b /dev/zvol/rdsk/rp1/iscsi/boot/esx1 ss1-esx1-boot

Most blog examples of enabling targets show the ZFS command line method as shareiscsi=on. This works well for a new iqn but if you want to allocate additional LUN under that iqn then you need to use this –b backing store method.

Now that we have some targets you should be able to list them using:

iscsitadm list target

Notice that we only see one iqn for ss1-zrp1, you can use the –v option to show all the LUN’s if required.

Target: ss1-zrp1
iSCSI Name: iqn.1986-03.com.sun:02:eb9c3683-9b2d-ccf4-8ae0-85c7432f3ef6.ss1-zrp1
Connections: 2
Target: ss1-zmp1
iSCSI Name: iqn.1986-03.com.sun:02:36fd5688-7521-42bc-b65e-9f777e8bfbe6.ss1-zmp1
Connections: 2
Target: ss1-esx1-boot
iSCSI Name: iqn.1986-03.com.sun:02:d1ecaed7-459a-e4b1-a875-b4d5df72de40.ss1-esx1-boot
Connections: 2

It would be prudent to create some target initiator entries to allow authorization control of what initiator iqn’s can connect to a particular target.
This is an important step. It will create the ability to use CHAP or at least only allow named iqn’s to connect to that target. iSNS also provides a similar service.

iscsitadm create initiator –iqn iqn.2000-04.com.qlogic:qla4050c.esx1.1 esx1.1
iscsitadm create initiator –iqn iqn.2000-04.com.qlogic:qla4050c.esx1.2 esx1.2

Now we can assign these initiators to a target and then the target will only accept those initiators. You can also add CHAP authentication as well, but that’s beyond the scope of this blog.

iscsitadm modify target –acl esx1.1 ss1-esx1-boot
iscsitadm modify target –acl esx1.2 ss1-esx1-boot
iscsitadm modify target –acl esx1.1 ss1-zrp1
iscsitadm modify target –acl esx1.2 ss1-zrp1
iscsitadm modify target –acl esx1.1 ss1-zmp1
iscsitadm modify target –acl esx1.2 ss1-zmp1

In order to boot from the target LUN we need to configure the QLA4050C boot feature. You must do this from the ESX host using the ctrl Q sequence during the boot cycle. It is simply a matter of entering the primary boot target IP set the mode to manual and enter the iqn exactly as it was listed from the iscsitadm list targets command. e.g.

iqn.1986-03.com.sun:02:d1ecaed7-459a-e4b1-a875-b4d5df72de40.ss1-esx1-boot

Once the iqn is entered the ESX host software can be installed and configured.
Till next time….

Site Contents: © 2008  Mike La Spina