Performance in Intel IOMeter DataBase Pattern
With the help of a DataBase pattern we measured the HDD performance during processing of requests with different queue depth and different share of writes among them. During the test session we used 8KB data blocks with random address.
To make it easier to analyze and discuss the obtained results we built diagrams for queue equal to 1, 16 and 256 requests. First of all we will consider the controllers performance in case of RAID 0 array:
As always let’s start with the queue depth=1. The graph above indicates that VIA VT8237 is the most efficient of all the testing participants. It outperforms the opponents under all types of workload. It only yielded to SiS964 in RandomWrite mode. Both Intel controllers get kind of defeated by their Taiwanese competitors. However, the newer ICH6R controller is faster than its ICH5R predecessor, especially when it comes to processing write requests.
The increase in the requests queue depth up to 16 appears more efficient for Intel’s controllers. ICH6R retains leadership until the 80% of writes. The second Intel controller appears better than SiS and VIA solutions until we reach 50% writes share. The performance of VIA VT8237 and SiS964 grows up tangibly as the share of writes increases, but the first controller of the two still looks somewhat better.
In case we have a requests queue of maximum depth, Intel controllers strengthen their positions. The indisputable winner here is ICH6R, which yielded only to SiS964 in case of all writes. ICH5R runs somewhat slower than its newer fellow and starts losing to it more noticeably when the writes share grows as big as 70%. The above mentioned SiS964 gets ever more efficient as the share of writes grows up, so that in the end it appears the winner. The VIA VT8237 solution turns o9ut kind of weak here, although its lag behind SiS964 is noticeable only when the share of write requests hits 70%.