HighPoint RocketRAID 2340

By Craig Simms, ZDNet.com.au on 20 May 2009 03:58 PM

Tags: highpoint, rocketraid, 2340, hba, pci express, cbs, drive, craig

It seems these days it's all about the storage — while the average consumer is turning to internet-based services (especially in part due to the explosion of high-definition digital media), the more tech-savvy are turning to either NAS or building their own file servers.

Thing is, how do you manage a lot of storage? With a thumping great controller card, that's how.

The four Mini-SAS ports of the controller, fed by the Marvell 88SX chips.
(Credit: Craig Simms/CBS Interactive)

The hardware
At the budget end for a 16-port controller card, HighPoint's RocketRAID 2340 is a PCI-E x8 card. The big chip hidden under the heatsink is an Intel QG41210 SL8DS Serial to Parallel PCI bridge (PDF), which sends signals from the PCI-E interface into two Marvell 88SX6081 (PDF) chips, which in turn drive two Mini-SAS ports each. Four one-metre Mini-SAS breakout cables are provided in the box, which spider out into another four SAS/SATA connectors each, allowing for a total of 16 drives.

The Intel QG41210 SL8DS translates from PCI-E to the Marvell chips.
(Credit: Craig Simms/CBS Interactive)

An SST 39VF020 (PDF) chip, hiding under a HighPoint sticker, holds the flashable firmware; on the rear is an Actel ProASIC Plus APA075 (PDF) ASIC, which HighPoint was unable to elaborate on what its purpose was. HighPoint's own HPT601LF is the XOR chipset, and for something that is the heart of the card, is surprisingly small.

The SST chip sits to the right of the HPT601LF and holds the BIOS. Further to the right is one of two Marvell 88SX6081s.
(Credit: Craig Simms/CBS Interactive)

The RocketRAID 2340 isn't entirely hardware RAID — it's what's known as an IOC, or I/O controller. This means the CPU will have to do some lifting, but the HPT601LF assists and offloads where it can under RAID 5 set ups, unlike an IOP (I/O Processor) set up where all the calculations are done on-card. Running things as IOC has a few benefits — compared to IOP it's cheaper, and compared to pure software RAID it should be faster thanks to taxing the CPU less. No battery backup and no cache further highlights the 2340 as an entry-level 16-port controller. As such, you won't ever find one of these in a mission critical server environment — it's better suited to your power user who is setting up their own file server.

The HPT601LF assists with RAID 5 calculations.
(Credit: Craig Simms/CBS Interactive)

You may have noticed HighPoint's phenomenal operating system support — the RocketRAID 2340 has to be one of the most supported expansion cards we've seen. Sadly, Solaris was not in the mix — while the OS supports the Marvell 88SX chip, the HPT601LF gets in the way. ZFS lovers will need to look to FreeBSD to get their kicks.

BIOS, web, GUI tools, oh my!
HighPoint offers three ways to manage its card — at the BIOS level, through a web interface or a GUI management tool. Linux users get a fourth in the form of a command line tool. It's worth noting that you can't have the web and GUI management tool installed simultaneously in Windows, or conflicts will arise and chances are you won't be able to contact your card.

The GUI tool for administration of the card really needs a graphical and usability overhaul.
(Credit: Craig Simms/CBS Interactive)

Upon first log-in through the management console you're prompted for a username and password, after which it asks you to change the password. This is client-based, and you can uninstall the software, reinstall and quite happily use the default password again. Doing it this way is not massively secure, but it's better than losing an IT guy who has the only password to the card.

The management console sadly looks out of date — Windows 95, perhaps even 3.1 era — and isn't the most well designed piece of software. You can't select multiple drives to then add to an array, for example; you have to add them one by one, a tedious task when you have 16 drives.

Managing the drive at the device level.
(Credit: Craig Simms/CBS Interactive)

You can manage on both the device and the array level, toggling such individual drive features as Read Ahead, Write Cache or NCQ, as well as viewing SMART status. Hot spare management is available, either as a general pool or assigned to a dedicated array. You can set the card to send email warnings should an event happen (good for jumping on a failed drive as quickly as possible), schedule tasks, and add users to the console management tool. The card is accessed by IP too, so it's possible to connect remotely to the device to manage it.

On the Linux side, things look much the same, just with the skin of your chosen desktop manager. GUI and web utilities were supplied as RPMs; while alien would convert them to .deb for Ubuntu, not everything would convert successfully, leaving the installed applications to complain that the driver wasn't installed. We soon discovered that the CLI utility included a .deb package, and if this was installed first, followed by the alien converted GUI installer, the GUI tool would work.

The Linux GUI tool looks much the same, just with GNOME skinning. Here we're managing the card at the array level.
(Credit: Craig Simms/CBS Interactive)

How we tested
Gathering up 16 one-terabyte drives is no small feat, and unfortunately this meant we ended up with a motley crew of brands and types to test the 2340 with. Those that take data integrity reasonably seriously though are likely to end up with drives from different batches anyway, so perhaps this is a more "real world" approach than to throw 16 Intel X25-E SSDs on it and see what happens.

*Driver 15.25 **Driver 182.50 ***BIOS 1.4, Driver 1.5 and GUI management tool 3.14-3

Setting up the array, volume limitations
We initially set up the array in the BIOS, adding all 16 drives to a RAID 5 array. This equated to 27.5 hours of background processing (or "backgroud", as the BIOS said). This was reduced to four hours, 19 minutes using the GUI management tool, as it supported foreground processing.

We quickly discovered Windows XP 32-bit doesn't support volume sizes larger than 2TB thanks to MBR — either Windows XP 64-bit or above is needed for larger sizes, as supported by GUID Partition Tables (GPT). Shortly after we were up on Windows Server 2008 64-bit, and the array was initialised for a total available space of 13,971.44GB under NTFS.

Ubuntu 8.04, as the officially supported distro on HighPoint's website, lead us to more weirdness, once the supported driver was up and running. After updating parted to 1.8.8 and creating our partition, we discovered mkfs.ext3 would only support up to 8TB partitions (odd, considering its quoted 32TB limit) — and so we settled on XFS instead, to the same capacity.

Finding the space, let alone power connectors for a 16-drive rig, is a daunting task without a rack mount case. (Credit: Craig Simms/CBS Interactive)

Benchmark tools are limited, too
The problems continue to benchmarks too. HD Tune Pro only benchmarks physical disks and so software RAID volumes don't appear — only their constituent disks. C't magazine's h2benchw, widely considered the most consistent benchmark tool available for Windows, will only test up to 2TB-sized volumes. HdTach is no longer actively developed and even its developers admit shortcomings. Iometer only realistically covers multi-user scenarios, something arguably the RocketRAID 2340 shouldn't be used for. Even Iozone on Ubuntu would hang on an mdadm array with test file sizes above 4GB, although performed just fine when RAID was handled by the card itself. To that end we employed CrystalDiskMark 2.2 for Windows, and the valid results from Iozone on Linux.

It's important to note that for our tests, we set the controller to use Write Back caching, not Write Through. This greatly increases write scores, but comes at the risk of losing data if the card is midway through writing and the machine loses power.

Performance
Initial boot up of the card can be quite slow, however, this can be altered by flashing the firmware and setting the load delay from anywhere up to eight seconds. The wait is purely to ensure that all the hard drives have finished spinning up before the card does its device check, and so you can tune based on the drives in your array.

In the same vein, turning on staggered spin up makes the start-up wait incredibly long, but only for a first time power-up — restarts, of course, don't need to spin up the drives at all, they're already running. If you've got a power supply though that can't handle the power draw of all the attached drives starting simultaneously, this is utterly necessary.

Throughput performance
In actual use, RAID 5 performance is where it's at, especially when you're talking about hardware-assisted controllers. We tested using both hardware-assisted RAID, and using Windows Server 2008's software RAID 5 (by setting each drive to its own JBOD array in the BIOS) to see what effect the HPT601LF had on performance. The following benchmark uses CrystalDiskMark 2.2 with the transfer size set to 1GB, and the number of tests set to five.

Software RAID gives higher reads, but hardware assisted gives vastly faster writes.
(Credit: Craig Simms/CBS Interactive)

Interestingly, Windows Server 2008 RAID 5 scored higher read scores (a good 65.5MBps higher on sequential reads). Presumably in this case the HPT601LF is incurring some overhead. This deficit though is nothing compared to the massive gulf on sequential writes that the HPT601LF has in its advantage (194.5MBps faster on sequential writes).

In an attempt to saturate the card, we also loaded all the drives into RAID 0 (14,902.88GB capcity) on HPT601LF, and included the results on the graph as a control. It managed an impressive 1046MBps sequential read, and 563.8MBps sequential write. Since it's based on PCI-E 8x with a maximum unidirectional bandwidth of 2GBps, we could clearly exploit this further with faster drives such as SSDs.

Let's take a look at Linux Iozone performance on XFS. We broke the tests down into 4K, 8K and 64K record sizes, and attempted to transfer file sizes up to 8GB to overcome the caching effects of our 2GB RAM.

Unfortunately, Iozone hung repeatedly while running tests 4GB and above on our mdadm array (while also giving unusable random read/write graphs), and so we were only able to pull read and write tests.

The graph below compares HighPoint's HPT601LF assisted RAID 5 versus Linux's software mdadm.

Write speeds slowly increase as file sizes increase. The hardware-assisted chip clearly shows its worth over software RAID through mdadm in writes.
(Credit: Craig Simms/CBS Interactive)

Write speeds for the HPT601LF clearly dominate the mdadm array, but only once file sizes exceed 32MB. You'll notice that the HPT601LF uses a little more CPU time than mdadm, but the payoff is quite substantial once you hit the higher file sizes.

Let's take a look at read speeds. Here we've had to throw away all the results before the 2GB point, as the caching effect of file and operating systems massively distort the results, giving average reads of over 4GBps — twice what the bus speed is capable of, and obviously incorrect. Unfortunately, due to this plus Iozone hanging for greater than 4GB benchmarks, this creates a rather unexciting graph.

Once the 2GB RAM threshhold is passed, read speeds level out.
(Credit: Craig Simms/CBS Interactive)

Looks like the HPT601LF assists read speeds better over Linux software RAID than Windows. CPU utilisation has been omitted due to inconsistent results through mdadm testing.

Conclusion
The HighPoint RocketRAID 2340 is a magnificent card for those wanting to run a cheap file server. It'll never end up in a mission critical file server thanks to its high CPU usage and lack of proper XOR, but at this price that's irrelevant. If you're looking for an affordable way to hook as many drives up as possible, plus have a bit of hardware RAID acceleration, it's hard to pass this one up.

• Related stories

Be the first to comment on this article!

• Just In

• Most Popular