About the Author

James is regular guest contributor to TechHead and is a Kent based qualified accountant, currently working in information security and technical architecture with  most of his  time “being spent on virtualisation and business continuity at the moment”. Check out his new virtualisation and storage blog here for more interesting and informative posts.

_________________________________________________

 The Dell EqualLogic PS4000 series of iSCSI storage arrays are positioned above the MD3000i series and targeted at the SME sector, especially for vmware virtualisation, as well as “branch office” use for larger companies.  The model reviewed here is the XV model, with sixteen 300GB 15k SAS drives, and dual controllers and dual power supplies.

This is quite a complicated product, so I’ve split this into four sections:

Part 1 – The PS4000

Part 2 – EqualLogic Networking with Force10

Part 3 – System Management and Monitoring

Part 4 – Performance

Part 4 – Performance

Performance

OK the important bit – the numbers that EqualLogic don’t seem to want to print.  I’ve tested this array in Raid-50, 5 and 6 using a Windows Server 2003 VM running IOMeter on a Dell R610 (with 1 vCPU and 4GB RAM allocated) using the methods described on my blog, here.  For comparison, I’ve also run the tests on a few different storage systems, including a single SATA drive and a few local RAID arrays (note that the IOPS graphs have a logarithmic scale).

Sequential Read

iSCSI with Gigabit Ethernet (GbE) just doesn’t provide great sequential throughput as GbE is good ‘only’ for about 110MB/s.  With some command-line configuration, as described in part 2, a pair of GbE’s can be used load balanced, providing total bandwidth between ESX and the EqualLogic of about 220MB/s.

Running in Raid-50, the PS4000 gets very close to saturating this network connectivity, providing 210MB/s with 64 outstanding IOs.  Raid-5 and 6 fair slightly less well, at about 170MB/s.  Still, compare this to the old Dell Clarion based AX150i iSCSI chassis, providing only 40MB/s:

Of course the throughput is dwarfed by the local Perc-6i with 5 SAS drives, which puts in an impressive 330MB/s.

Sequential Write

The first configuration tested was Raid-6, which provides about 120MB/s on writes.  Given the dual-parity nature of Raid-6, this seems like reasonable performance.

But then there is a surprise – Raid-50 and Raid-5 provide similar performance at 130 and 140MB/s respectively.  Again for comparison, the local Perc-6i sustains over 330MB/s in the same test using Raid-5.

Still, sequential throughput isn’t a deal breaker and is relatively unimportant for virtualised workloads expect perhaps for volume based backups, and even then these figures are likely to be easily sufficient for perhaps 90% of real-life workloads.

Random Throughput

Random IO is what this array is all about – serving multiple contending workloads without allowing latency to rise enough to impact application response times.  This is where the PS4000 really scores over local storage, primarily due to its 16 drive bays.  With 3.4ms average seek and 2ms latency, each 15k6 should give at least 185 IOPS.  In practice,  command queuing and the fact the test file is very much smaller than the overall capacity, in an effort to simulate real life workloads, mean numbers considerably higher than this can be expected in these tests.  EqualLogic pre-sales presentations claim around 270 IOPS per drive in RAID-5 – but does it deliver?

The random workload used is 8K blocks, 100% random IO, 70% read, against a 30GB test volume.  The Raid-6 performance seems amazingly strong, just 10% down on Raid-5 although this is essentially due to the 70:30 weighting of the testing.

Achieving these IOPS rates needs a good queue depth – performance levelling out at about 64 outstanding IOs.  At this level, and even up to 192 outstanding IOs, latency remains comfortably within vmware’s recommended 50ms.

Recovery Mode

After a disk failure, the array immediately starts a rebuild using the spare drive.  As a result of this, performance is impacted – the IOMeter results roughly halved whilst this was ongoing.  Sequential write was hit hard, being reduced to around 30MB/s.

In Summary

The EqualLogic PS4000 is an accessible storage system for the SME market built to the usual Dell standards which performs well, scales now up to fourteen units, and should prove reliable thanks to top-quality components throughout.

Failure of disks, power supplies, interconnecting switches and controllers seems to work well and is gladly demonstrated by the commissioning engineer.  Dual-controllers also enable firmware updates to be made without downtime (although a maintenance window should still be used).

Limitations

The system isn’t however without it’s limitations.

In testing, an old Oracle 8i import job proved a corner-case where the array performed badly.  It should be noted that this is a pretty ancient piece of software that works with a low queue depth and is unlikely to see much use today – the problem essentially being the underlying command latency, which will always be higher for iSCSI compared to local storage.

Another limitation is present in preparing for a remote replication deployment.  The system works with block change tracking utilising (internally) 64MB blocks – hence a database changing a single sector will trigger a 64MB transfer to the remote site.  It would be helpful if the system could be used with the block change tracking enabled without a replication partner, in order to collect the statistics needed to assess the viability of SAN layer replication prior to committing a good wedge of cash to it.

Another issue is with the thin provisioning, however this is another iSCSI limitation.  Industry seems to be demanding iSCSI over NFS despite the protocols being very similar in terms of efficiency.  The issue with iSCSI though is that there is no delete command, so a run-away log file in a VM running on a thin-provisioned volume, or a carelessly thick-provisioned VM, will wipe out those gigabytes on the EqualLogic pretty much forever.  The only way to retrieve the space from the LUN is to move the VMs to a new LUN and delete the old.  In contrast, NFS storage simply doesn’t suffer part of this issue since ESX is communicating with it at the file system level (zeroing free space and a storage vMotion would still be needed to shrink thin-provisioned VMDKs).

Finally the sequential write performance seems disappointing, but it should be noted that the interoperable PS6000 series have higher speed processors and 10-GbE in some models, and would probably be better suited to such workloads.  In random performance – much more of interest in virtualisation projects and in particular database workloads – the EqualLogic is of course way ahead of a typical 6-drive server, running as it does at over 4,000 8K IOPS in Raid-50 with latency of under 16ms with 64 outstanding IOs.

Tuning

Getting the network layer configuration right is the key to getting the best from these arrays.  The mostly browser or Java based management interfaces offer excellent performance monitoring capabilities, but there is almost nothing by way of tuning.  This is really a good thing though since the system is self managing and will automatically move data between arrays based on their performance profiles in use.  And less settings to adjust really means less to get wrong.

vStorage

As the long awaited vStorage API seems finally due to make an appearance, the EqualLogic product will get even better since it will be one of the first able to provide hardware off-load capabilities to ESX, such as snapshots.

And Finally

After the release of the first part of this series several months ago, Dell were quick to get in touch to offer any assistance needed including direct access to a high-level engineer and access to an invitation-only user group.  The testing methodology has been reviewed by them and the performance metrics provided here were confirmed to be in the range expected.

The few niggles aside, the EqualLogic PS4000 is an excellent product with excellent management tools, and its support backing and simple configuration (networking aside) are real strengths in this product.