mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Hardware > Cloud Computing

Reply
 
Thread Tools
Old 2017-12-31, 23:22   #1
ATH
Einyen
 
ATH's Avatar
 
Dec 2003
Denmark

5×683 Posts
Default Price per core

Right now c4.large is not the cheapest per core anymore at least in Ohio. c4.8xlarge is actually cheapest right now, but also because it is very cheap in 2b compared to 2a and 2c, and because the c4.large have gone up a lot this month.

c4.large: 1 core (2 vCPUs): $0.0205: $0.0205/core
c4.xlarge: 2 cores (4 vCPUs): $0.0388: $0.0194/core
c4.2xlarge: 4 cores (8 vCPUs): $0.0789: $0.0197/core
c4.4xlarge: 8 cores (16 vCPUs): $0.1565: $0.0196/core
c4.8xlarge: 18 cores (36 vCPUs): $0.3255: $0.0181/core

Last fiddled with by ATH on 2017-12-31 at 23:23
ATH is online now   Reply With Quote
Old 2017-12-31, 23:39   #2
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

2×32×653 Posts
Default

Andreas, can you also post analogous c5 prices? I'm curious how the c5.9xlarge you said was best for your new-prime-verify run using Mlucas_avx512 compares vs the c4.
ewmayer is offline   Reply With Quote
Old 2018-01-01, 00:34   #3
ATH
Einyen
 
ATH's Avatar
 
Dec 2003
Denmark

341510 Posts
Default

Ok. There is no c5 in Ohio, so will use NVirginia, it is about the same price as Oregon, while Ireland is more expensive. Only those 3 has c5 for now.

Btw the new year is "c4" (2018)

c5.large: 1 core (2 vCPUs): $0.0301: $0.0301/core
c5.xlarge: 2 cores (4 vCPUs): $0.0602: $0.0301/core
c5.2xlarge: 4 cores (8 vCPUs): $0.1222: $0,0306/core
c5.4xlarge: 8 cores (16 vCPUs): $0.2407: $0,0301/core
c5.9xlarge: 18 cores (36 vCPUs): $0.5416: $0.0301/core
c5.18xlarge: 36 cores (72 vCPUs): $1.0832: $0.0301/core

So almost the same price per core for all of them, and roughly 50% more expensive than the c4's.

Last fiddled with by ATH on 2018-01-01 at 00:48
ATH is online now   Reply With Quote
Old 2018-01-01, 01:03   #4
Mark Rose
 
Mark Rose's Avatar
 
"/X\(‘-‘)/X\"
Jan 2013

25×3×31 Posts
Default

Quote:
Originally Posted by ATH View Post
it is very cheap in 2b compared to 2a and 2c
Keep in mind each account has a different map of virtual availability zone to physical availability zone. So anyone doing a price comparison won't necessarily see the same pricing tied to the same zones.
Mark Rose is offline   Reply With Quote
Old 2018-01-01, 01:57   #5
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

2×32×653 Posts
Default

Quote:
Originally Posted by ATH View Post
Btw the new year is "c4" (2018)
You must be on a different calendar, because I get 7E2. :puzzled:

Quote:
So almost the same price per core for all of them, and roughly 50% more expensive than the c4's.
So running decently-tuned avx512 code will make the c5 at least breakeven in term of flops/$, and possibly a few % cheaper. Remind me - the c4 you got ~4 ms/iter on in the new-prime verify was which type? We need to compare flops/$ for that vs the 3 ms/iter you got on the c5.9xlarge.

Of course for 'production' LL testing, the 1-and-2-core instances of c4 and c5 would be cheaper per flop than the manycore ones, so that comparative math remains to be done.
ewmayer is offline   Reply With Quote
Old 2018-01-01, 01:59   #6
GP2
 
GP2's Avatar
 
Sep 2003

13·199 Posts
Default

But price per core is not the whole story.

You are still better off running 18 c4.large (1-core) instances or 9 c4.xlarge (2-core) instances rather than one c4.8xlarge (18-core), at least for mprime, because performance doesn't scale linearly with multiple cores on the same physical machine.

In any case, it's a pretty unusual situation pricewise right now, since the c4.large and c4.xlarge have risen about 50% in just the last month. This is probably due in part to the usual end-of-quarter and holiday e-commerce reasons, which will soon cease to be a factor. However, it is probably also due to the new smoother spot pricing model and the new spot instance interruption options of "stop" or "hibernate", in addition to "terminate", which make spot instances a lot easier to use. This causes more business users to shift their applications away from on-demand and reserved instances and toward spot instances, and this demand might mean sustained higher spot prices.
GP2 is offline   Reply With Quote
Old 2018-01-01, 02:03   #7
GP2
 
GP2's Avatar
 
Sep 2003

50338 Posts
Default

Quote:
Originally Posted by ewmayer View Post
You must be on a different calendar, because I get 7E2. :puzzled:
Maybe he's using base 53.
GP2 is offline   Reply With Quote
Old 2018-01-01, 03:09   #8
ATH
Einyen
 
ATH's Avatar
 
Dec 2003
Denmark

5×683 Posts
Default

Quote:
Originally Posted by ewmayer View Post
So running decently-tuned avx512 code will make the c5 at least breakeven in term of flops/$, and possibly a few % cheaper. Remind me - the c4 you got ~4 ms/iter on in the new-prime verify was which type? We need to compare flops/$ for that vs the 3 ms/iter you got on the c5.9xlarge.
4.42 msec/iter = 226.2 iter/sec on c4.18xlarge for $0.3255: $0,00144 per iter/sec.
vs
3.03 msec/iter = 330,0 iter/sec on c5.9xlarge for $0.5416: $0.00164 per iter/sec

Last fiddled with by ATH on 2018-01-01 at 03:14
ATH is online now   Reply With Quote
Old 2018-01-01, 05:33   #9
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

2·32·653 Posts
Default

Quote:
Originally Posted by ATH View Post
4.42 msec/iter = 226.2 iter/sec on c4.18xlarge for $0.3255: $0,00144 per iter/sec.
vs
3.03 msec/iter = 330,0 iter/sec on c5.9xlarge for $0.5416: $0.00164 per iter/sec
Thanks, but I think - correct me if I'm wrong - you intended this, given that instance prices are per-hour:

4.42 msec/iter = 226.2 iter/sec on c4.18xlarge for $0.3255/hr: $0.000000400 per iter.
vs
3.03 msec/iter = 330,0 iter/sec on c5.9xlarge for $0.5416/hr: $0.000000456 per iter

Thus both net out for between $30-40 for the new-prime verify run @4608K FFT length. On the c5 we get ~2.4 msec/iter @4096K, so running at that length drops the run cost to under $30, still slightly higher than the $20 Batalov mentioned for his EC2 verify run of the previous new prime. The higher costs likely reflect the EOY price hike GP2 mentioned.

Last fiddled with by ewmayer on 2018-01-01 at 05:34
ewmayer is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
new I9 line price/spec, upto 36 core firejuggler Hardware 14 2017-09-26 05:28
Price is Right Race Game c10ck3r Puzzles 19 2013-03-27 14:00
On the price of electricity in Denmark fivemack Lounge 1 2011-01-23 19:54
price cuts moo Hardware 3 2006-11-21 17:49
Processor Performance vs. Price louis_net Hardware 9 2004-06-25 17:13

All times are UTC. The time now is 12:19.


Fri Dec 2 12:19:00 UTC 2022 up 106 days, 9:47, 0 users, load averages: 0.74, 0.84, 0.83

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2022, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.

≠ ± ∓ ÷ × · − √ ‰ ⊗ ⊕ ⊖ ⊘ ⊙ ≤ ≥ ≦ ≧ ≨ ≩ ≺ ≻ ≼ ≽ ⊏ ⊐ ⊑ ⊒ ² ³ °
∠ ∟ ° ≅ ~ ‖ ⟂ ⫛
≡ ≜ ≈ ∝ ∞ ≪ ≫ ⌊⌋ ⌈⌉ ∘ ∏ ∐ ∑ ∧ ∨ ∩ ∪ ⨀ ⊕ ⊗ 𝖕 𝖖 𝖗 ⊲ ⊳
∅ ∖ ∁ ↦ ↣ ∩ ∪ ⊆ ⊂ ⊄ ⊊ ⊇ ⊃ ⊅ ⊋ ⊖ ∈ ∉ ∋ ∌ ℕ ℤ ℚ ℝ ℂ ℵ ℶ ℷ ℸ 𝓟
¬ ∨ ∧ ⊕ → ← ⇒ ⇐ ⇔ ∀ ∃ ∄ ∴ ∵ ⊤ ⊥ ⊢ ⊨ ⫤ ⊣ … ⋯ ⋮ ⋰ ⋱
∫ ∬ ∭ ∮ ∯ ∰ ∇ ∆ δ ∂ ℱ ℒ ℓ
𝛢𝛼 𝛣𝛽 𝛤𝛾 𝛥𝛿 𝛦𝜀𝜖 𝛧𝜁 𝛨𝜂 𝛩𝜃𝜗 𝛪𝜄 𝛫𝜅 𝛬𝜆 𝛭𝜇 𝛮𝜈 𝛯𝜉 𝛰𝜊 𝛱𝜋 𝛲𝜌 𝛴𝜎𝜍 𝛵𝜏 𝛶𝜐 𝛷𝜙𝜑 𝛸𝜒 𝛹𝜓 𝛺𝜔