Streaming Patterns
After we’ve checked out the basic features of the electronics of our HDDs, we can discuss the second type of load – the dependence of speed on the number of simultaneous data streams.
In the diagrams below the results are sorted by the averaged read/write speed for each of the four cases. That is, the models that slow down the most when processing multiple simultaneous streams are at the bottom of the diagram. The numbers show the HDD’s maximum speed of reading/writing a single continuous data stream (in 64MB blocks).
The two drives from Samsung have the highest results (we already noted their record-breaking areal density in the descriptive part of this review). The third-generation DiamondMax 10 is close behind the leaders. Maxtors can also be seen at the bottom of the diagram – the same model but with non-typical 83GB platters. Interestingly, the newer Seagate Barracuda 7200.9 is some 2MB/s slower than the Barracuda 7200.8.
When the HDDs have to read data alternately (for example, from two logical volumes), they form three groups with very different results. The Maxtor DiamondMax 10 didn’t lose anything in speed. All the drives from Western Digital and the Hitachi T7K250 lost half their speed while the “old” Maxtors and the Hitachi T7K250 together with all the HDDs from Seagate slowed down tenfold! The HDDs from the third group are not suited for paralleled load as their look-ahead reading algorithms are intended for one data stream only. You can see that all the manufacturers, except Seagate, took this kind of load into consideration in their newer products, but Maxtor’s engineers did just a fantastic job. Samsung’s HDDs, formally belonging to the same P120 generation, perform in an interesting way: the SP2514N obviously has multi-threaded look-ahead reading (like its predecessors), but the SP2504C with a Serial ATA interface lacks it for some reason. This must be the consequence of the attempt to introduce Native Command Queuing into its firmware.
The picture is different when we test the more realistic situation in which the requests are queued. Most of the HDDs nearly reach their maximum speed by reading one data stream mostly, paying little attention to the second stream. The six HDDs with support for Native Command Queuing behave differently, but four of them couldn’t maintain a high speed for both of the streams. The first-revision Maxtor DiamondMax 10 and its counterpart from the MaXLine III series are the only drives to cope with the task brilliantly: they provide equal performance to both consumers (thanks to NCQ) and don’t slow down (thanks to unique look-ahead reading algorithms). The second and third revisions of the same model show a different behavior, reading the first stream to the detriment of the second stream while maintaining almost the same total speed. Can it be that the NCQ support is disabled in them? We’ll check this out in the next section. Right now let’s examine the dependence of write speed on the number of data streams.
It’s already not surprising to see the Maxtor DiamondMax 10 in the lead but the last places of both generations of Hitachi drives are somewhat unexpected. The best the Deskstar T7K250 could do was to overtake the worst one of the Seagate HDDs. Interestingly, the Barracuda 7200.8 with enabled NCQ is almost as fast as the leaders while the two drives from Samsung are equally good at writing despite their different behavior. It is clear that the firmware algorithms are the main factor for this test. Next goes the amount of cache memory, and only then – the linear read speed.
