Dedicated machine for benchmarks

[mcollina]

Intel is donating a dedicate server to run benchmarks, and nearForm is operating that machine from our HQ.
The machine is built on the Kaby Lake architecture (I'll update on specs).

We need to hook this up to the ci system.

cc @jasnell @piccoloaiutante

Ref: nodejs/node#8157

[gibfahn]

Do you know what OS it's running? We'll need to make sure it's covered by our existing ansible scripts.

nearForm is operating that machine from our HQ.

What does that mean exactly? Intel are paying for it but it's your hardware?

[mcollina]

The HW is Intel's, we are paying for the operating expenses and hosting.

[mhdawson]

In terms of hooking it up to the CI, for other machines the approach is that you provide the root password to the build WG who will then manage the machine. In terms of things like reboots etc all existing machines either have some sort of control panel that we can use to reboot when necessary. I'm guessing that won't be the case in this instance so we'll need contacts/procedures to be able recover in the cases were the build team members can no longer access the system.

We'll need to discuss based on the system/specs but unless there is some type of virtualization that we believe will not affect performance I'm assuming they will be a single OS instance running on each of the boxes.

[jbergstroem]

Better understanding who has physical access (and intended os level access) would also make it easier for us to understand how the machine can be used.

[mcollina]

Our staff has physical access, but you can refer to me for any physical maintenance. I can forward account credentials for a root-level account.
You can refer to me for any physical maintenance, I'll forward the request to our team.

As it is a dedicated server, it should be used for benchmarks.

[jbergstroem]

@mcollina benchmarks of course, it's just mostly about what code we can benchmark so to speak. For instance, adding it to a release process where we benchmark prior release to avoid any regressions could include things that comes from security-related release stuff.

[mcollina]

@jbergstroem that's ok. Also check benchmark regressions on PRs, maybe?

[rvagg]

Fixed @ #800

Temporary use is simple test running, more sophisticated benchmarking use is left as an exercise for the ... someone else. Let me know if there's anything I can help with in advancing that cause though.

[mcollina]

cc @AndreasMadsen

[AndreasMadsen]

@mcollina anything specifically I will comment on?

Primarily, I wanted a machine so we could run a Jenkins job like that proposed in nodejs/benchmarking#58. The biggest problem with the benchmarks right now, are that the benchmarks take too long to run on a personal machine and developers tends to use it for other things while it's benchmarking. The latter creates systematic noise that no statistics can't handle.

Running the entire benchmark suite on each release (maybe just major/minor) sounds nice. However, it will take a very long time, we are talking multiple days. Also, when running a single benchmark test there is a small risk of a false positive, this is not a problem when running just a few benchmarks but when running them all (1000+) there will be many false positives, thus it doesn't have much value. There is no theoretical nice way around this, the only thing that can really be done is to make the benchmarks run longer. This is one of the reasons why the Large Hadron Collider costs 7+ billion dollars :p

[octaviansoldea]

Following some previous communication with colleagues, I would like to indicate that the Node.js
benchmark servers provided by Intel have the following details:

The machines are Wildcat Pass 2U 8x3.5" HDD 10Gbe Xeon DP v4 Server [R2308WTTYS-IDD], 1x1100W,
equipped with Intel® Xeon® Processor E5-2600 v4 Family, Socket R3, Intel® Server Board S2600WTTR,
Board Chipset: Intel® C612 Chipset. They are equipped with 8 slots of memory where each one is of type 8GB 2400 Reg ECC 1.2V DDR4 Kingston KVR24R17S8/8I Single Rank.
Moreover, each station is equipped with integrated LAN 2x 10GbE, the number of LAN Ports is 2, and has hard disk of 1 TB, SATA, 6Gb/s, 7200 RPM. This type of server aims Cloud/Datacenter.

[mhdawson]

@AndreasMadsen I think nodejs/benchmarking#58 is the first thing we should aim to get running on these machines. Once we get experience with that we can see if we think regular runs (like on releases) might make sense.

[mhdawson]

I'll see if @gareth-ellis has time to get together in the next few weeks to see if we can push nodejs/benchmarking#58 forward.

[rvagg]

@octaviansoldea thanks for the specs! I might link to your comment from the nodes on Jenkins.