[HH] $ per core for desktop parts

[Kurt Keville]

Since VM clouds are mostly a discussion of core count, I suspect cache
sizes and SIMD extensions are not fair to bring into the conversation
unless you get those for free (e.g. NEON with ARMv7). Still you would
have to be evangelizing a pretty synthetic benchmark to make a case
for a smaller core count VM cloud.

I have one of those needs right now; I need to have a distcc cloud for
a distro repository backend. the more virtual cores / threads the
better. So 2 anemic processors are probably better than 1 good one. We
need it mostly for the RV128 port of Debian (soon to be at
http://riscv.mit.edu/ ) Right now we are doing it entirely with
RISCVEMU and we will likely have to do it again when binutils breaks
again (not scheduled yet, but it is imminent).

On Sun, May 21, 2017 at 9:18 AM, Shankar Viswanathan
<shankar.viswan at gmail.com> wrote:
> I always find it hard to understand what most folks mean by "VM
> performance". A VM is just a container, and unless you are benchmarking the
> OS running inside it, "VM performance" by itself doesn't indicate much.
>
> Personally, what I care about is the application(s) being run inside each
> VM. Is that application single or multi-threaded, CPU or IO or memory bound?
> Different hypervisors and different CPUs could have all kinds of
> optimizations underneath, but it all boils down to whether your application
> can take advantage of those features to get near native performance. While
> $/thread is a good metric, the nature of your application will determine if
> that is the best metric - e.g. there are certain applications where enabling
> SMT actually hurts performance.
>
> Just my $.02 as a chip designer,
> Shankar
>
>
>
> On 05/20/2017 03:03 PM, Tom Metro wrote:
>>
>> As AMD has released some relatively high-core count desktop parts for
>> low cost, I've been wondering how well they might do for building a VM
>> host. I haven't researched yet to see what people are saying. I seem to
>> recall that their prior FX line didn't perform as well as one might
>> expect for VM loads, given their core count.
>>
>> Micro Center sent out an email ad this week featuring a bunch of AMD and
>> Intel parts, and out of curiosity I took a look to see how they compared
>> based on cost per core.
>>
>> $ per Cores/$ per thread (CPU cost)
>> AMD   FX 8320E $10/$?  ($80)
>> AMD   FX 8350  $14/$14 ($110)
>> Ryzen  5 1600  $33/$17 ($200)
>> Ryzen  5 1600X $38/$19 ($230)
>> Ryzen  7 1700  $40/$20 ($320)
>> Ryzen  5 1400  $43/$21 ($170)
>> Intel i5-7500  $48/$48 ($190)
>> Intel i5-6500  $45/$45 ($180)
>> Ryzen  5 1500X $48/$24 ($190)
>> Intel i5-7600K $50/$50 ($200)
>> Intel i7-6800K $55/$28 ($330)
>> Ryzen  7 1800X $58/$29 ($460)
>> Intel i5-6600K $50/$50 ($200)
>> Intel i3-7350K $65/$33 ($130)
>> Intel i7-6700K $70/$35 ($280)
>> Intel i7-6850K $75/$38 ($450)
>>
>>
>> So the AMD FX parts win on this metric with a mere $10 to $14 per core,
>> but they're antiques. They don't support hyperthreading, and the $10
>> part's specs didn't even list thread count. On a per-thread basis, the
>> Ryzen parts are nearly as cheap. Unclear why anyone would buy one today.
>> (Though I could see an unscrupulous OEM stuffing them in a budget
>> machine.) They don't even seem well suited to appliance use, like NASs
>> or media players, as they have too many cores for that. Yet Micro Center
>> keeps pushing them, well after the release of Ryzen parts.
>>
>> The next best are the low-end Ryzen 5 series parts, attaining $33 per
>> core. In comparison, low-end Intel parts are no bargain, as they top out
>> at 4 cores, and often have no hyperthreading.
>>
>> As the price goes up on these desktop-optimized parts, the per-core
>> speed goes up, but the core count doesn't scale proportionally. No
>> surprise there, as many games and apps still respond best to higher
>> clock on fewer cores. The $450 i7-6850K ranks worse of all these at
>> $75/core, while the corresponding $460 Ryzen 7 1800X attains $58/core
>> due to having 2 more cores than the Intel part (6 vs. 8 cores).
>>
>>  From an overall density perspective, the Ryzen 7 1700 at $40/core is
>> probably the best bet, as it packs 8 cores into one part, with 16
>> threads. You can perhaps slice that up into 8 VMs with 2 virtual cores
>> each.
>>
>> So if one wanted to build up a low-budget cloud, populating with Ryzen 5
>> parts (avoid the 1600X as it uses 95W compared to the non-X version's
>> 65W, and is only slightly faster) or Ryzen 7 1700 (also 65W) might be
>> the way to go. (The later gains you 2 cores for $120. That density
>> premium might be worth it even for a small scale build from a pure cost
>> basis. Every 3 nodes built with the Ryzen 7 gains you the equivalent of
>> an additional node built with the Ryzen 5, but costs you $360 more.
>> Unlikely you can build a Ryzen 5 node for less than that (CPU $200, MB
>> $100+, RAM $200+, etc.). Though given how non-server motherboards are
>> capped at 32 GB RAM, the lower core density approach would be better if
>> you have high RAM workloads, or even require high storage bandwidth.)
>>
>> Hopefully we'll see some server parts from AMD soon with even better
>> $/core, and lower power. There also seems to be almost no MiniITX boards
>> available for Ryzen so far. (I see one Asrock board that's sold out.)
>> And MiniITX would be preferred for building a high-density cluster.
>>
>>   -Tom
>> _______________________________________________
>> Hardwarehacking mailing list
>> Hardwarehacking at blu.org
>> http://lists.blu.org/mailman/listinfo/hardwarehacking
>
>
> _______________________________________________
> Hardwarehacking mailing list
> Hardwarehacking at blu.org
> http://lists.blu.org/mailman/listinfo/hardwarehacking