Optimizing Performance, Like Vertex 4
Performance Versus Capacity
Nothing surprising happens if we read sequentially across these drives at a queue depth of one. The Barefoot-based SSDs don't perform as well as the Plextor drive, but that's not a big deal. Frankly, this isn't what we were looking for anyway; we need to switch to writes.
Now this is more like it. Plextor's M5 Pro does exactly what we expect it to. But both OCZ drives lose a ton of performance after 50% of their capacity is filled. We see a few transient drops prior to that point, but past the halfway point, throughput looks a lot bleaker. The 200,000 KB/s we're showing is equivalent to 195 MB/s, and neither drive even sustains that. As it turns out, OCZ is using the same kind of SLC emulation mode we identified in OCZ Vertex 4 128 GB: Revisiting Write Performance With Firmware 1.5. When there isn't much data being processed, the information fits into the pseudo-SLC envelope. But once the drive is hammered, it has to swap some of what it has stored around. When that data is on the drive for long enough, it eventually figures out where to put everything. But in the interim, performance is certainly impacted.
We don't know much about how this process works, but it's clearly an attempt to work smarter, rather than harder. Without it, the Vertex 450 wouldn't have the front-end speed to put it at the top of the stack. It's reasonable to assume that this fits in to OCZ's desire to round out its platform with more premium SSDs able to command higher prices. Slugging it out at the bottom-end is tough when your competition manufactures flash as well. So, focusing on performance-oriented SSDs does seem like a better long-term strategy.
Performance Versus Time
In concept, this is similar to the charts we just presented. We're just plotting write speed over time, or how long it takes to sequentially write to the entire accessible LBA space. Plextor isn't as fast to start, but it sustains higher average throughput. That means its M5 Pro finishes first, taking 560 seconds to write from the first LBA to the last. OCZ's Vertex 450 and Vector are a different story. They spend just a fraction of their time writing more than 500 MB/s. The other three-quarters occurs south of the 200 MB/s mark. In other words, the Vertex 450 takes an additional 400 seconds to write the allocated capacity once.
But what about a second consecutive fill?
The second fill occurs immediately after the first; there's no time for background garbage collection or any other downtime optimization. Neither OCZ drive takes as long to finish this task, averaging higher performance than the fill preceding it. Since the entire accessible space is already full at this point, all three drives are writing almost exclusively over used blocks.
OCZ's Vector puts some distance on the Vertex 450, but doesn't break away completely. Despite a messier-looking data series, the two OCZ drives complete the second fill faster, while the M5 Pro's run is unchanged. We can see the Vector and Vertex 450 clawing to get back to a better performance state. Specifically, the Vertex 450 spends half of the run approaching 350 MB/s. But once it does, it stays there. The Vector is similar, hugging 400 MB/s after grinding its way back up.
Now, here's the thing: during normal usage (even in taxing workloads), you're going to be hard-pressed to spot this behavior. Other companies are using SLC NAND emulation (notably SanDisk), but more as multi-tiered caching approach. OCZ appears to implement its strategy at a global level, less granularly. Without more information from the vendor, it's difficult to comment definitively. But it's fair to say the company's optimization does work, particularly in our Storage Bench trace on the next page.
