Messages in this thread | | | Date | Sat, 23 Oct 1999 09:35:14 -0400 | From | Douglas Gilbert <> | Subject | Re: ultra2 scsi support? |
| |
Gerard Roudier wrote: > > On Fri, 22 Oct 1999, Douglas Gilbert wrote: > > > On a related theme, I have an Advansys 940UW controller > > connected to a IBM DCHS04U which should allow for 40 MB/sec > > of bandwidth. But /proc/scsi/advansys/0 reports: > > .. > > 0: Transfer Period Factor: 12 (20.8 Mhz), REQ/ACK Offset: 8 > ???????? > > Transfer period factor 12 -> period 50 nano-seconds which can be reported > as _exactly_ 20 MHz frequency even if that does not make full sense given > that the period is some minimal delay requirement in SCSI. > > > Furthermore with tests using the SCSI READ BUFFER command I > > can measure >30 MB/sec throughput. My guess is that this > > ultra, u2 and u160 is done at a lower level and hidden > > from the messaging level lest older things get upset. > > The data throughput does not make sense if you donnot consider the command > overhead. This overhead is not negligible at all when using ultra-2 and > will be an important differentiation point with Ultra-3, IMO. > > Given that the average IO chunk is not that large in real applications, it > is important that softwares, controllers and devices result in a very low > command overhead. Bursting max data throughput in data phase is far easier > to achieve and I generally assume devices are quite able to sustain such > bursting throughput. > > Example of calculation: > > I can get up to 8900 TPS, reading sequentially 2 very fast drives at a > time in Ultra-2 mode using 4KB actual IO chunks (on a 53C896). Result is: > > - Total command latency = 1000000/8900 = 112 micro-seconds per > transaction. > - Data phase latency _assuming_ 80 MB/s data phase throughput = > 12.5 * 4 = 50 micro-seconds. > > Result is that command overhead is less than 112-50 = 62 micro-seconds. > > As a rule of thumb, speaking of throughput/bandwidth without speaking of > _latency_, when speaking about transactions based devices is chatting in > the wind.
For more "chatting in the wind" see http://www.torque.net/sg [follow "timings" link]
The test first finds the maximum size buffer supported by the SCSI READ BUFFER command and then times a 200 MB transfer based on that buffer size. Yes, it is contrived.
It is not trying to measure total data throughput from the magnetic media to the user space. It is trying to time the SCSI bus transfer speed and adapter DMA speed in a Linux environment. It was done to shed some light on whether direct IO (i.e. DMAing directly into the user space) was worth the effort.
Back to my orginal point. The measurement demonstrated that the SCSI bus throughput was greater than what the Advansys driver was reporting while still being consistent with the hardware being used.
Doug Gilbert
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |