Single-drive Iometer performance
IOMeter will allow us to put some stress on these chipsets' implementations of Native Command Queuing. By queuing up read and write commands and reordering them opportunistically, NCQ is able to minimize some of the effects of the relatively high head seek times (at least in computer terms) of hard disk drives. We tested the nForce4 SLI Intel Edition with and without NCQ enabled.

We also tested the AMD version of nForce4 with two different drivers, the older one, 5.07, released in February, and the new one, 5.18, just recently released. There's some history here, as well. We noticed a problem with NVIDIA's disk controller some time ago, in our chipset RAID comparison, when the nForce3 250Gb's controller wouldn't go any higher than 128 I/Os per second with single drive in IOMeter. The controller also hit a wall, though at a higher transaction rate, with multi-drive RAID arrays. This problem persisted in our nForce4 Ultra review last fall. We noted at that time that the problem wasn't a showstopper, but it's not a good performance characteristic for a disk I/O subsystem not to scale properly.

It appears NVIDIA has fixed this problem in its newer drivers on both the Intel and AMD versions of the nForce4, as you can see below.

We don't have numbers for the nForce4 chipsets at queue depths of 256, by the way, because IOMeter crashes on the nForce4 at this queue depth, as it does on several other types of storage controllers. Fortunately, NVIDIA has looked into this problem and identified a bug having to do with threading and memory allocation and proposed a patch to correct it.

You'll also note that command queuing doesn't appear to kick in with either of the nForce4 chipsets at load levels below 16. NVIDIA says they don't issue queued commands at low queue depths because doing so adds unwanted overhead and negatively impacts performance. That assessment is consistent with some of our testing, which has shown lower performance in desktop applications with NCQ enabled. NCQ will be a godsend with lots of outstanding requests, though, as the IOMeter results indicate.

All in all, this is an impressive performance by the nForce4's disk controller, at least in a single-drive configuration. The nForce4's CPU utilization is ever so slightly higher than the Intel controller's, but the differences are pretty minor. We'll have to test RAID performance, including RAID 5, in a future article.