John Robinson put forth on 1/3/2011 10:00 AM: > On 03/01/2011 06:41, Stan Hoeppner wrote: >> http://www.addonics.com/products/host_controller/ad5sapm.asp >> http://www.buy.com/prod/addonics-ad5sapm-serial-ata-controller-5-x-7-pin-serial-ata-300-serial/q/loc/101/213272437.html >> If your mobo SATA ports support FIS based switching, this PMP will give >> you 5 SATA II drive ports. It doesn't use a PCI slot of any kind. No >> additional software required. No kernel driver issues. 300MB/s is >> sufficient for 5 drives in an mdraid setup isn't it? > For a backup array, yes, but I'm not sure it is for online storage. > 300MB/s is an absolute max and there's protocol overhead etc, but even > if it's minimal we're still looking at no better than 50MB/s per drive, > while the drives can manage 125MB/s these days. When using PMPs one will always have less theoretical b/w per drive than what that drive can push on paper with a streaming read. Considering that over 90% of real world workloads are random IO heavy, not streaming, it's unlikely you'll ever run out of b/w using a PMP based setup. I haven't. Also, as has been pointed out, you will be limited by PCIe x1 2.0 b/w before you are limited by qty 2 SATA II ports: 500 MB/s for PCIe x1 rev2.0 versus 600MB/s for 2xSATA II links. This is true whether you use a 4 port PCIe rev 2.0 card with 4 direct attached SATA II drives or 10 drives with 2 PMPs. You only have one PCIe slot. To get more bandwidth, you'll have to use an x4 or x8 card, which is way above your stated price range. > I doubt my motherboard supports FIS PMPs. It's an Asus P5Q Pro, Intel > P45+ICH10R, and I'm pretty sure the ICH10R doesn't support PMPs even if > the original spec said it would. I just looked it up, and the ICH10R does _NOT_ support PMP. Neither does the onboard Silicon Image 5723 chip, which actually connects to the southbridge via a SATA II port (stupid). BTW, the P5Q Pro has 3 PCIe x1 slots and 2 PCIe x16/x8 slots, plus 2 PCI slots. You told us the only slot you have available is 1 PCIe x16/x8. What is consuming the 3 PCIe x1 slots? The PCIe x1 slot just North of the top x16 slot should be free. I'm guessing the other 2 are blocked by your GPU cooler, correct? Picture: http://www.asus.com/product.aspx?P_ID=qH6ZSEJ8EPY6HoNU&templete=2 Move your GPU card to the bottom x16 slot and you free up all 3 x1 slots. This gives you a lot more options to get the solution you want inexpensively. > There is a Marvell 88SE6121 SATA+IDE chip on there but it's currently in > IDE-only mode for the DVD drive and even if I switched over to SATA mode > and a SATA DVD drive that'd only give me one more SATA port. But it > might work with a FIS PMP, I suppose. Originally I was going to propose a $20 Sil 3132 based 2 port card for use with 2 of the Addonics PMPs, but all those I could find are PCIe rev 1 only, for only 250MB/s of b/w. All of the SI SATA host controller chips work perfectly with Silicon Image PMPs (3276 chip) and the kernel drivers have no issues. If you had two PCIe x1 slots available, using two such cards with one PMP plugged into each would work very well indeed. This is what I do. Your b/w would max at 500MB/s, but again, this is more than sufficient for the vast majority of workloads. Keep in mind that this bandwidth is bidirectional, so you actually have 1GB/s of throughput with reads and writes in flight simultaneously with multi-user or multi-threaded workloads, especially with a deep NCQ buffer. > I'd do that too - no problems doing case mods here. I suppose it's > possible the mounting holes might be able to be made to line up with > some of the mounting holes on the side of the hot-swap chassis. On the > other hand I might cheat and use the little plastic mounts with > double-sided tape on their feet. Yeah, there are lots of possibilities for mounting these handy little PMPs. I've seen folks just put that really thick double sided foam tape on the back side and stick them to an open spot on the bottom or side of the chassis. > [...] >> The driver for the Marvell chip is present in kernel >> 2.6.19 and later. Considering that 2.6.19 is like 6 years old, I'd hope >> your kernel is newer. > > It's kernel-2.6.18-194.26.1.el5 so it's stuffed full of backports and > security updates, it's less than two months old. Yes, I have sata_mv, > but several people have reported data corruption issues with some > Marvell controllers - a bad interaction with SMART I think. I recommended the Marvell based card strictly because it does PCIe x1 rev 2 for 500MB/s. I only use the Silicon Image based cards, but I use more than one if I have more than 4 drives in an array. The cost comes out the same at $40 as the Marvell card, but the SI route requires 2 PCIe slots, and you only have one. >> It may be a little more money than you were planning on spending, but >> for little more than the cost of one hard drive > > In this case I'm using consumer-level drives so they're about Â40 ($60), > so $165 is a bit rich for me, especially since it's potentially limited > for throughput. Well, you've given us conflicting requirements that are impossible to meet with any solution: 1. Low cost 2. Full b/w per drive 3. Only one PCIe slot availiable There is not a solution available to meet all of your criteria. Either you must spend more on the card to get all the SATA ports and b/w you want, or you have to go the PMP route to lower costs, and sacrifice some theoretical performance. Below is probably the closest you will get to what you want but the price is higher than what you've said you want to spend. It'll give you 8 direct attach SAS/SATA II ports with a PCIe x4 interface: http://www.newegg.com/Product/Product.aspx?Item=N82E16816115057 $149 But you'll have to buy 2 fanout cables: http://www.satacables.com/sata_multilane_SAS_cables.html FAN-OUT-20INCH4X Which run $20 apiece, $40 total. So you're up to just under $200. Your most cost effective option for adding a 5 drive cage, by far, is this: http://www.newegg.com/Product/Product.aspx?Item=N82E16816124032 But it's only PCIe x1 at 250MB/s. > Nevertheless, thank you very much for taking the time for such a > considered reply. You're welcome. I think you're finding yourself in that "I want to have my cake, but eat it too" situation. You can't get what you want at the price point you want. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
