Navigate: Advice Beginners BIOS Guide CPUs Links Mainboards Memory Network Storage Video/Sound Cards Contact Forum SiteMap Sponsors Home PC Hardware from Made-in-China.com Search: Enter your search terms   Web www.lostcircuits.com Submit search form	.
Prices:

LOSTCIRCUITS		SHORTCUTS: Accepting a Narrow Bus Internal vs. Interface Performance HDTach: ATA vs. SATA HDTach vs. WinBench99 vs. PCMark Parallel ATA Command Overhead Streamlined SATA FIS, Prelude to Performance WinBench99 Overview Winbench99 Details Business Winstone2002, File Copy Conclusions Cheetahs and Cudas on the web Comments?
Seagate Barracuda SATA-V     Kudos to Cudas
(Review by MS, February 4)

One of the new buzzwords is First Party DMA (FPDMA) of SATA, which is often thrown in and never explained, even though it is rather simple: Compared to the non-deterministic role of Parallel ATA devices, Serial ATA drives can decide by themselves which data to transfer and when. That is, any transfer starts with the host sending a Read FPDMAQueued command after which it disconnects from the bus. A SATA drive is capable of actively reconnecting to the bus, meaning that it does not have to wait for the polling by controller when the latter "just feels like it is time". Because the SATA drive itself is deterministic, it can issue a DMA Setup Frame Information Structure (FIS). This FIS contains the information about the nature of the data and the memory location the data are supposed to go into, which was included in the FPDMAQueued(Ext) FIS originally issued by the host. To make a long story short, SATA uses a streamlined protocol without the 100% command overhead in queued Parallel ATA or else the bus occupancy in immediate service mode.

How Does it Translate Into Performance?

Theory is theory and practice is practice. In addition, there are driver issues to be taken into consideration that can optimize performance while cutting down on CPU usage, especially with respect to the IDE interface. Likewise, prefetching algorithms can be applied to enhance IDE performance by speculatively loading data into the drive cache. However, it is possible to make certain predictions regarding the outcome of the various benchmarks. That is, any benchmark that either relies on the internal drive performance or else measures data transfers by looking at large chunks like those typical for content creation applications will not show much of a difference between equivalent drives using only different interfaces.

Differences will become overt, in workloads consisting of random accesses with transfers of small files that can burst from the cache. In this case, the streamlined command overhead will contribute to the overall performance. Examples are mostly business applications, data base accesses and similar.

WinBench99 2.0 scores using the onboard Silicon Image RAID controller in single Barracuda SATA V configuration. Scores shown are Business Disk Winmark99 (blue) and High-End Disk Winmark 99 (red), depending on which drivers were used and whether HyperThreading was enabled. The drivers in question are the standard "Base" drivers that are complemented by some new SiI Filter drivers. The data shown here are averages of a minimum of 5 runs and meant only as a preamble to illustrate the complexity of what we are dealing with. Results with HT enabled and base drivers only were lower than those shown with HT disabled but there is no sense in beating a dead horse. Bottomline is that the devil is in the drivers that can be used to show a performance increase of 100% in certain applications. Else, the omission of these drivers will have a strong negative impact on the outcome of the benchmarks. There is more to that story, though, as we will show in the following.

next page:    => nForce2 Platform =>