 |
 |
Response in-line... Martin. -- Martin Bly RAL Tier1 Fabric Manager > -----Original Message----- > From: Testbed Support for GridPP member institutes [mailto:TB- > [log in to unmask]] On Behalf Of Ben Waugh > Sent: Saturday, June 26, 2010 10:06 AM > To: [log in to unmask] > Subject: Advice on procuring worker nodes > > Dear All, > > This is an appeal for guidance from both those who have put out > tenders for CPU nodes, and those with knowledge of what makes a good > worker node for ATLAS in particular. > > I know procedures vary between institutions, but I have been advised > by our Procurement department to do a "mini tender" involving the five > suppliers who have framework agreements to supply servers to UCL, > asking for the greatest possible CPU power for a fixed price. Generally a good strategy if you know what you want to specify and can do so in clear terms. > The HEPSPEC rating is the obvious measure to maximise, but not all > suppliers have the means or inclination to run a specialised benchmark > for a relatively small order, about £40k. How have others done this? We are generally asking for more than £40k-work of kit but we specify a total performance requirement by HEP-SPEC06, but impose several other constraints (power use limits per rack for example). > Do you restrict yourselves to the suppliers who already have > experience in dealing with GridPP and can run HEPSPEC themselves, or > do you use other benchmarks or some less direct way of comparing the > CPU rating of the products on offer? For EU-level tenders (200kEuro+?) we can't restrict who responds to the 1st stage which is a trawl for suppliers. We evaluate these and then pick several to issue the ITT to. For non-EU-level tenders, there are various local levels most of which result in some form of competitive bidding. Restricting in general is a useful in some senses in that you can for example approach suppliers familiar with your requirements. But this may limit your access to best price from a supplier who isn't on your list. Most suppliers you are likely to deal with should be able to get the SPEC06 benchmark and run it given a few weeks. Most of our suppliers have been able to turn numbers around in the limited time we give them and any who have responded to Tier1 tenders any time in the last year or so should be familiar with this test by now. Generating your own performance metric would be a hard task unless you keep it very simple - and that would be unlikely to match the performance seen in real use. One thing to note: two systems that are specified the same and look the same but have different motherboards and memory configurations but the same CPUs *are* likely to give slightly different HS06 scores so it isn't a good idea to invest too much faith in a solution from supplier A being the same performance as an ostensibly identical solution from supplier B. I've seen 2-5% differences. > There are of course other factors affecting job throughput, including > hard disks and RAM. Is there some way of measuring the effect of > these, or would you just set a minimum requirement on both and then > maximise the HEPSPEC? We just use HS06 and the hardware cost plus total power for 5 years to generate a 'TC05' cost to adjudicate against - this encourages greener solutions at slightly increased hardware cost. > If you would take the latter approach, what is a > sensible trade-off between disk performance and price? Presumably > 10kRPM SAS disks will be better than 7.5kRPM SATA, but maybe a striped > pair of slow disks would be an alternative? And how much disk space do > you allow per CPU core? Disk(s): 100GB per core + 100GB for system. We haven't in the past tried to be clever with disks so we have tended to ask for a single 7200rpm SATA disk. This year we may decide to ask for multiple disks to try and future proof against disk i/o load. RAM: 3GB/core + option for an upgrade - usually cheaper than a retro-fit. > If there is anything else I haven't asked but you think I should > consider, please tell me that too! Maintenance offerings. Can you get spares outside your maintenance period. Running SL5 (particularly NIC drivers!). > Best regards, > Ben > > -- > Dr Ben Waugh Tel. +44 (0)20 7679 7223 > Dept of Physics and Astronomy Internal: 37223 > University College London > London WC1E 6BT