On Thu, Apr 01, 2010 at 05:38:32PM +0100, Ed W wrote: > I have been hanging around the gluster mailing list recently and one of > the things you see with network filesystems is people not noticing how > much the latency is going to kill you. eg if your latency was 1ms to I'm extremely aware of latency. Which is why I talk SSD and us, not 7200 rpm HD and ms. > the iSCSI server then under some circumstances you will max out at 1,000 > IOs per second. Now depending how that translates to app performance A single spindle SATA with 1 TByte has 75 IOPS, while an consumer MLC SSD has up to 9 kIOPS for random 4 k reads, up to 25 kIOPS random 4 k writes. Even if this is mostly marketing, that's an almost three orders of magnitude difference. > this can be a huge performance hit and presumably why Adrian commented > that even 10gbit FC with it's massively lower latency can still limit > bandwidth to each client to a fraction of the total server bandwidth. Again, I'd rather use a 200 EUR SSD within the guest host rather than an external iSCSI target. It's just lacking native hybrid zfs support I would have to do it manually. It's probably just a stupid idea, and I should stick two 80 GByte SSDs in RAID 0 (doubles the write endurance, at only slightly increased background failure rate) and leave it at that. Mount everything else via NFS. Expensive, but effective. > I think your question is so academic as to be much better to simply > benchmark some tests and literally see what will work best for you? In I asked because I don't have the hardware yet, and assumed somebody else had benchmarked this or has this in production. At that matter, I do not have a good workload model, since I don't have hundreds of loaded guests in production, just idling ones. As things go, a simple nmap will make the firewall almost crap out. > general an SSD will help most if you have io wait problems and it's less > beneficial for streaming reads/writes? I don't worry about streams. Arguably, SSD (especially RAID 0 SATA SSD) are very good at streams. With low end Atoms, I'll be most likely CPU-bound. > A datapoint though, but I recently upgraded my Macbook Pro to 6GB from > 4GB ram and also bought a 256MB SSD to replace the spinning disk it had > before. Now the problem I was trying to solve was that I was short of > free memory while running some Windows 7 instances under parallels and > the machine was literally freezing completely for seconds at a time > while switching apps. Now the SSD arrived first and simply changing > this *completely* transformed the machine, literally night and day, I've been using a 80 GByte SSD in an Atom netbook as well as RAID 10 stripes over 8x 160 GB consumer Intel MLC SSDs for read-mostly Oracle. No official benchmarks, but I'm completely sold. Not a single failure yet, which is a lot more than I can say about SATA HDs (Seagate <cough> 7200.11 <cough>). > whole machine is lighting fast and zero slowdown booting some win7 > instances even with low ram. I then added the extra ram a day later and > noticed no appreciable difference in performance. Now this is > COMPLETELY the opposite way around to what I expected, but interesting - > usually I would add ram first and faster disks second. What I worry about is flash wear with small MLC SSDs. If the swap sees a lot of use, it's toast pretty soon. So is write-intensive databases. > However, this is a desktop and here the speed of opening an application > (which only happens a dozen times a *day*) is the benchmark by which I > judge the speed of the computer! In contrast on a server we want to > optimise for steady state IOs happening throughout the day and it's less > clear how the SSD will benefit things here... I'm thinking a dramatic enhancement, but I was wondering about experience of people actually doing it. > So I think you need to benchmark your setup and perhaps look instead at > something like a controller with writeback cache rather than an > SSD/iscsi? I see you can get the Dell PERC things quite cheap on ebay > for example? Again, I'm working with 300 EUR server kits, 90 EUR memory which take less than 20 W without spinning bits. > Just a thought... > > (P.S. Anyone with a laptop running slowly, the SSD blows the doors off > with perceived performance - seriously impressive...) -- Eugen* Leitl <a href="leitlhttp://leitl.org">leitl</a> http://leitl.org ______________________________________________________________ ICBM: 48.07100, 11.36820 http://www.ativel.com http://postbiota.org 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE