Re: [gentoo-project] Replacement hardware planning for Gentoo VM hosting

[Matthew Thode]

On 05/19/2016 06:47 PM, Robin H. Johnson wrote:
> Infra has already discussed most of this hardware planning in
> #gentoo-infra, but I thought it might be useful to see any other
> comments on the hardware plan. If you wish to make private comments to
> this thread, please send directly to infra@g.o or
> gentoo-core@l.g.o instead of the gentoo-project list.
>
> Remarks like 'you should use ZFS instead of this' aren't directly
> helpful to this discussion. What is more useful is pointing out any
> potential problems you might see with the plan, or gotchas in the
> hardware.
>
> We've previously run Ganeti [0] with general success, and we'd like to
> continue doing so (vs libvirt or openstack). It offers VM storage
> redundancy via DRBD (amongst other options), which we're going to take
> best advantage of by using a cross-over 10Gbit link between two nodes
> (as we have no 10GBit switching in the environment). Some of the VMs
> will run on spinning disk, others on SSD, others maybe w/ dm-cache.
> libvirt IS an easy fallback from Ganeti, but lacks some of the automated
> failover and DRBD handling options.
>
> This will house at least the following existing VMs, all of which have
> large storage needs:
> - woodpecker.gentoo.org
> - roverlay.dev.g.o
> - tinderbox.amd64.dev.g.o
> - devbox.amd64.dev.g.o
>
> And virtualize the following older systems:
> [2007 Dells]
> - finch.g.o (puppet)
> - vulture.g.o (GSoC host)
> [2010 Atoms]
> - bellbird.g.o (infra services)
> - bittern.g.o (blogs webhost)
> - bobolink.g.o (rsync.g.o node, dns slave)
> - brambling.g.o (bouncer, devmanual, infra-status)
> [Other]
> - meadowlark.g.o (infra services)
>
> And New VMs/services:
> - split git to rsync & snapshot generation from dipper?
> - split blogs (and other) database hosting from dipper?
>
> We'd probably keep the two other 2011 Dell systems in operation for the
> moment, to distribute load better, but have enough capacity to run their
> VMs as when they fail.
>
> The general best prices we've seen are from a vendor that's new to us,
> WiredZone, and we're willing to give them a try unless somebody has even
> better pricing to offer us.
>
> Hardware (all in $USD):
> Supermicro SYS-2028TP-DECTR [1][2]
> - $2,732.42/ea, quantity 1
> - two half-width 2U nodes in a single chassis w/ shared redundant PSU.
> - each node has:
> - 2x 10GBe ports (there are no SFP options)
> - 12x 2.5" SAS3, controller in JBOD/IT mode
> Per node:
> Intel Xeon E5-2620v4 [3] -
> - $421.56/ea, quantity 2
> 32GB DDR4 PC4-19200 (2400MHz) 288-pin RDIMM ECC Registered [4],
> - $162.89/ea, quantity 4
> - require min of two DIMMs per CPU
> - price jump to 64GB DIMMs very high.
> - buy more RAM later?
> Seagate 2TB SAS 12Gb/s 7200RPM 2.5in, ST2000NX0273 [5]
> - $315.18/ea, quantity 4
> - 4-disk RAID5 (mdadm)
> Samsung 850 EVO 1TB, MZ-75E1T0B/AM [6]
> - $345.00/ea, quantity 2
> - RAID1 (mdadm)
> = $3445.40/node
>
> Overall cost:
> $2,732.42 - chassis
> $3,445.40 - left node components
> $3,445.40 - right node components
> $  315.18 - 1x spare ST2000NX0273 HDD
> $   25.00 - 3ft CAT6a patch cable (estimated)
>
> Parts sub-total: $9,963.40
> Labour sub-total: $300 (estimate)
> Taxes: $0.00 (Oregon has no sales taxes)
> S&H: $200 (estimate)
>
> Grant total: $10,463.40 (USD)
>
> Future hardware improvement options:
> - Add more RAM
> - Add up to 6x more disks per node.
>
> [0] http://www.ganeti.org/
> [1] http://www.supermicro.com/products/system/2U/2028/SYS-2028TP-DECTR.cfm
> [2] http://www.wiredzone.com/supermicro-multi-node-servers-twin-barebone-dual-cpu-2-node-sys-2028tp-dectr-10024389
> [3] https://www.wiredzone.com/intel-components-cpu-processors-server-bx80660e52620v4-10025960
> [4] https://www.wiredzone.com/supermicro-components-memory-ddr4-mem-dr432l-sl01-er24-10025993
> [5] https://www.wiredzone.com/seagate-components-hard-drives-enterprise-st2000nx0273-10024175
> [6] https://www.wiredzone.com/samsung-components-hard-drives-enterprise-mz-75e1t0b-am-10024043
>

+1 to this generally, the one question I have is if we want to spend ~1k
more on one of the 4x nodes, it'd allow us to expand easier in the
future.  The reasoning against this that I can think of is that we want
a higher disk/node ratio (which is just 6 per node instead of 12).

-- 
-- Matthew Thode (prometheanfire)