Date: Sun, 24 Aug 1997 04:42:44 -0700 (PDT) From: Simon Shapiro <Shimon@i-Connect.Net> To: freebsd-scsi@freebsd.org Subject: DPT - New Release Message-ID: <XFMail.970824044244.Shimon@i-Connect.Net>
next in thread | raw e-mail | index | archive | help
Hi Y'all, I have just posted on sendero-ppp version 1.2.3 for both 2.2 and 3.0 Some points: * Both patches are against the cvs archive as of Friday night. * The SMP behavior is unknown as of yet. It booted and mounted filesystems, so it is not totally dead :-) * The release passed regresion and initial stability test on two of our machines. * Performance testing was performed only with the debug configuration, not the production (DPT vs. SENDERO). In 2.2, with 64 processes hammering a 6 device, 2 busses RAID-0 array (24GB) atriped 8k, we saw a very stable 900 tps. This is with st running 4k read - modify - write. Load average was about 50. Increasing the load beyond 256 processes pushed down the io/sec rate. Most likely due to cache starvation (only 32MB). * We did not test performance on 3.0 yet. We are still trying to break it. The reason for these numbers is not to brag, nor start a heated thread. But rather to demonstrate what to expect. I have been in contact with several people who were concerned about what numbers to look for. As you noticed, I do not publish (nor care :-) abount MB/sec. We use these systems for heavy database work. We only care about the number of transactions per second. BTW, we have yet to see a DPT exceed 1,700 transactions per second per controller. Highlights: * Justin's CAM interrupts were integrated into both 2.2 and 3.0 and seem to work fine. If (in the context of our testing method) higher SPL numbers indicate better throughput, then it went up drastically. We see, in 3.0, on 256 processes (plus a forever loop catting /etc/termcap over telnet session) LA of about 90. This is with only two disks on one bus. * We re-worked the critical sections to give as good a perfromance as possible. Accepting commands from the upper layer and scheduling it can be done in parallel with interrupt service. As a whole, the system is as parallel as can be (pending Justin's agreement :-). * There are no more calls to splhigh anywhere in the driver. * We added the spldlm as well, so when we release the dlm code, it will just drop in. * All the code to support Unix dptmgr is now in the driver. It received very superficial testing (if any). Please try to download this version and torture it. We would like to have it finalized before 7-Sep-97. This is it for now. Thanx for your support. Simon
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?XFMail.970824044244.Shimon>