Server Infrastructure Upgrade Weekend - AMD EPYC Rome Across The Board

[phoronix]

Phoronix: Server Infrastructure Upgrade Weekend - AMD EPYC Rome Across The Board

There may be some brief, intermittent downtime this weekend as all of the Phoronix Media web server infrastructure is being upgraded. This also marks the first time in Phoronix's 16+ year history that the servers are AMD powered, thanks to the incredible success of the EPYC 7002 "Rome" series...

http://www.phoronix.com/scan.php?page=news_item&px=New-Servers-2020-Rome

[vladpetric]

Does DRAM speed matter that much when you have such beefy caches on the EPYC 7002? (A minimum of 32 MiB of L3 cache according to https://www.amd.com/system/files/doc...-Datasheet.pdf )

[Michael]

Does DRAM speed matter that much when you have such beefy caches on the EPYC 7002? (A minimum of 32 MiB of L3 cache according to https://www.amd.com/system/files/doc...-Datasheet.pdf )

It can make a difference depending upon the workload. When there isn't too much of a difference in pricing, sad it's not 3200 or at least 2933, but more was bummed by seeing just 2x32GB rather than quad channel configuration at least.

[vladpetric]

It can make a difference depending upon the workload. When there isn't too much of a difference in pricing, sad it's not 3200 or at least 2933, but more was bummed by seeing just 2x32GB rather than quad channel configuration at least.

Do you happen to have benchmark numbers where you explore such a difference in RAM speeds with processors that have huge L3s? Ideally, everything the same except for dram imbalance.

P.S. I really appreciate your general approach to benchmarking. I remember for instance you isolating IPC gains over the last decade or so, by controlling for the other factors (that was pretty awesome).

[swagg_boi]

These are exciting times, I shall light a firework for Phoronix's independence from imperial Intel tomorrow night

[Michael]

Do you happen to have benchmark numbers where you explore such a difference in RAM speeds with processors that have huge L3s? Ideally, everything the same except for dram imbalance.

P.S. I really appreciate your general approach to benchmarking. I remember for instance you isolating IPC gains over the last decade or so, by controlling for the other factors (that was pretty awesome).

I know I ran some benchmarks for Rome looking at memory speed impact, but can't find them on the site so maybe lost track of publishing them. If so after the migration is complete next week can run some new benchmarks.

[CochainComplex]

Good coice ...btw its also nice to see that subscribers money is spend*
wisely

..invluding german beer

[tildearrow]

Michael Yes, time to tell you.

Often loading node links (like the ones in emails) take like 15 seconds.

[Michael]

Just some current latencies:

$ ping openbenchmarking.org
PING openbenchmarking.org (162.252.83.89) 56(84) bytes of data.
64 bytes from 162-252-83-89.static.hvvc.us (162.252.83.89): icmp_seq=1 ttl=52 time=156 ms
64 bytes from 162-252-83-89.static.hvvc.us (162.252.83.89): icmp_seq=2 ttl=52 time=151 ms
64 bytes from 162-252-83-89.static.hvvc.us (162.252.83.89): icmp_seq=3 ttl=52 time=156 ms

Firefox network latency of http://openbenchmarking.org title page is 300ms.

Yes, hoping it will be better. Try Sunday

[Teggs]

It can make a difference depending upon the workload. When there isn't too much of a difference in pricing, sad it's not 3200 or at least 2933, but more was bummed by seeing just 2x32GB rather than quad channel configuration at least.

I suppose you have tried to convince them? You aren't just some random user after all, you have the exact data you're referencing and can run any further test they like.

In fact, they should be your customer, too.