Yeah, Allyn isn't a PR guy, but for Wendell he'll turn up with gear and geek out. This guy knows his stuff more than almost anyone else on the planet when it comes to storage.
Allyn is as well spoken as a great PR rep, talking about stuff with the depth of knowledge of a senior engineer at Intel. Serious props, this was great to watch.
Omg it's Allyn! I always loved his in depth knowledge about storage and memory and I remember this one story he told about being in a submarine and having to fix something before they got caught being somewhere they weren't supposed to be.
14:23 - In APL and J, we use memory-mapped I/O a lot. That lets us write clean, elegant algorithms that that have NO I/O, so we get superb performance and the mathematics is not hidden behind all that memory clutter. Several days of your basic database work turn into ten minutes of simple code that you can explain to your colleague when playing You Bet Your Company.
I'm gonna be straight with you Wendell. This may be the next level content on this channel. I'm getting some serious "Jay Leno's garage" vibes from this video. 1 You bring in an interesting computer 2 have a passionate conversation about it and get into the details 3 You say "Can we take it for a spin" and do something fun with it.
Back in the day when Allyn was still at PCPer and reviewed the first optane products, he helped me pick optane disks as swap for plain old kvm hypervisors. It basically made our memory pool for virtualization about 3x in usable size with overcommitment in runtime state. Imagine the savings comparing optane SSDs to the same capacity of RAM... please invite him more often!
The conversation about NUMA and cache migration/copying reminded me of the Extreme Black Magic involved in the SGI Altix systems I used to admin. Maintaining cache coherency across multiple rack cabinets is terrifying and amazing when you dive deep enough down the rabbit hole.
Would be interesting if the CPU had non volatile cache or a mirror of it so if you pull out the power cable and plug it back in it continues right where it left off in a millisecond.
it's impossible to have a perfect 1:1 non volatile cache, at least with current technology. L1 cache access is in the 1ns range, and registers are sub-nanosecond... they would've been written to a thousand times over before you could sync to optane pmem.
Maybe a stupid question regarding client SSDs with DRAM but no powerloss protection: If you power the SSDs not just with the computer‘s PSU but use small powerbanks for 5V and 12 V, would this add functioning powerloss protection?
Yeah, if someone breaks in, rips all of them cables and Optane P5800Xs out but leaves the Xeon platform completely alone I wouldn‘t know at all who might have done it…🙃
Pmem isn't a free lunch. CPU's have several levels of caches and write buffers that have the same basic role and semantics. Yes. When you write to pmem it doesn't actually get to the media unless you use the special flush instructions. So you end up doing the exact same crazy stuff done for filesystems for you pmem where you write unsynced and then flush at critical checkpoints. In that respect it's the same as even the client SSDs.
@@malventano First off, that was a fun video. Sorry if I came off like I suggesting it's not a solved problem, that wasn't my intent. My point is that it is solved, exactly like dealing with buffered drives is a solved problem. Same reasons for having buffers, causing the same problems, leading to the same solutions. I co-wrote a patent for a phase change memory in a DIMM in 2009 so I'm a fan of the concept. Nevertheless, I've long thought pmem advocates do a disservice to the cause by running with talking points that bury the lede. Pmem shines plenty in low latency and small operations, no reason to pick fights you can't easily win. Saying pmem is better because it doesn't buffer data, is just inaccurate and frankly unnecessary. Integration of pmem tech into SW stacks is hard, the ecosystem has made it more accessible, but the point is it can be well worth the effort.
I don't quite get the interest in the Samsung MZ-1LV9600 ... kinda old news (on a 2022 video) ..? But these guys know this subject better than I (by far) so I assume I'm who's missing something... but what..?
Around 2 years ago Linus dropped a 10k CPU. Today I have a system with two Intel Xeon platinum p-8124 cpu's. In four year if I'm not playing minecraft Java edition in Linux mint on equivalent to this system I'll be very disappointed in myself.
Does this break the Von Neumann architecture requiring different compilers as storage is really memory and data? Couldn't your data and instruction memory be the same thing? Just curious likely I am not understanding it.
Why do you need this kind of a server? It is useless for OLAP because 4TB is almost nothing in terms of disk space. It might be a valid platform choice for a search engine, in-memory queue or an OLTP workload, but why do you need these servers unless they are not so expensive? It feels like you still need checksums because even ECC memory fails when you have 100k machines.
@@malventano Why didn't you "travel by telecom rack"? Is PMem too delicate or were you afraid of ending up in LMG Studios? Referencing this BTW (When Wendell helped Gamers Nexus build a new NAS): ua-cam.com/video/Hix0l8cFaMw/v-deo.html (this is the punchline of a 2min joke)
But wait -- really .. ? In "10 years this won't be a couple grand on eBay" ..? I'm skeptical of that. Highly in fact. I think tha'ts exactly what it'll be by then. (at most)
Sadly, I think it's time to pull the plug on this channel. Between the recent video discussing the pope and pets to this one talking about hardware fantasies at a time when I cannot get mainstream memory, I don't see the relevance of the channel anymore. I'm not even sure who this channel is for anymore and I've been in IT since the 70s. Best of luck in the future.
If Allyn is from Intel, they need to send him more often! He was a very refreshing as a PR person that actually seems excited about the tech.
The glory days were when Allen would do deep dive reviews of drivers for PCPer
@@dylansmith9215 "deep dive," in this case, being insufficient to describe how obsessively detailed those pieces were.
Yeah, Allyn isn't a PR guy, but for Wendell he'll turn up with gear and geek out. This guy knows his stuff more than almost anyone else on the planet when it comes to storage.
i'm interested in allyn's official intel title.
bug tester?
@@acasccseea4434 Storage Technical Analyst.
I'll take two. Seems tasty.
this man cannot be stopped. DO NOT EAT THE SILICON.
Allyn is as well spoken as a great PR rep, talking about stuff with the depth of knowledge of a senior engineer at Intel. Serious props, this was great to watch.
Omg it's Allyn! I always loved his in depth knowledge about storage and memory and I remember this one story he told about being in a submarine and having to fix something before they got caught being somewhere they weren't supposed to be.
14:23 - In APL and J, we use memory-mapped I/O a lot. That lets us write clean, elegant algorithms that that have NO I/O, so we get superb performance and the mathematics is not hidden behind all that memory clutter. Several days of your basic database work turn into ten minutes of simple code that you can explain to your colleague when playing You Bet Your Company.
This is mad science of insanity. "I'm just gonna yeet 100GB in a second and see what happens. Oops the computer got scared acted weird."
I'm gonna be straight with you Wendell. This may be the next level content on this channel. I'm getting some serious "Jay Leno's garage" vibes from this video. 1 You bring in an interesting computer
2 have a passionate conversation about it and get into the details
3 You say "Can we take it for a spin" and do something fun with it.
I always love these nerd chats, even if they talk about stuff that is completely out of my reach.
Back in the day when Allyn was still at PCPer and reviewed the first optane products, he helped me pick optane disks as swap for plain old kvm hypervisors. It basically made our memory pool for virtualization about 3x in usable size with overcommitment in runtime state. Imagine the savings comparing optane SSDs to the same capacity of RAM...
please invite him more often!
Nice to see you Allyn, excellent stuff as always.
"Not worried to be losing data". Even with "persistent storage" we're doing RAID etc to provide redundancy. Are we going to do PRAM RAID now? :D
Always great to hear from Allyn. Thanks for bringing him on for a few videos.
I really liked this episode. Allyn was def a treat to watch as he appears to be genuine in his love for tech nerd deep dives.
Just awesome! ^^ Thanks Wendell and Allyn!
Finally! That is what I've been expecting to happen since optane was first announced! It's very exciting!
The conversation about NUMA and cache migration/copying reminded me of the Extreme Black Magic involved in the SGI Altix systems I used to admin. Maintaining cache coherency across multiple rack cabinets is terrifying and amazing when you dive deep enough down the rabbit hole.
Would be interesting if the CPU had non volatile cache or a mirror of it so if you pull out the power cable and plug it back in it continues right where it left off in a millisecond.
it's impossible to have a perfect 1:1 non volatile cache, at least with current technology. L1 cache access is in the 1ns range, and registers are sub-nanosecond... they would've been written to a thousand times over before you could sync to optane pmem.
that was HELLA INSIGHTFUL!!! I have never given much thought to computing on this level, and it really is bonkers!!! THANK YOU FOR THIS!!!
Allyn always comes baring gifts :) Great stream.
Another awesome storage chat. Eager for the next.
Tomorrow :)
We need more Allyn!
Great Stream, love Allyn..
i understand like 1% of the topic, but i do enjoy the conversation. keep em coming
Thanks Wendel, this is the good stuff!
That was brilliant. Thank you.
I saw Allyn in the thumbnail and had to click, for the PCPer nostalgia sound of his voice
Maybe a stupid question regarding client SSDs with DRAM but no powerloss protection: If you power the SSDs not just with the computer‘s PSU but use small powerbanks for 5V and 12 V, would this add functioning powerloss protection?
fun chat indeed. love a good nerdfest. hail to the king, baby.
STILL waitung on that "sound test" we were promised
Father Wendell. In the thumbnail I thought Wendell was wearing a priest collar.
And woo hoo, Columbus is getting an Intel fab!
Wendell going towards server rack...
...servers sweating profusely.
CES? What CES? This was just 🤗
Wendell, Allyn, Storage, I'm in!
RIP Optane: 2017 - 2022
Yeah, if someone breaks in, rips all of them cables and Optane P5800Xs out but leaves the Xeon platform completely alone I wouldn‘t know at all who might have done it…🙃
What a wholesome video
And I was excited about a new 1t gen 4 m.2
Awesome server and talk!
Pmem isn't a free lunch. CPU's have several levels of caches and write buffers that have the same basic role and semantics. Yes. When you write to pmem it doesn't actually get to the media unless you use the special flush instructions. So you end up doing the exact same crazy stuff done for filesystems for you pmem where you write unsynced and then flush at critical checkpoints.
In that respect it's the same as even the client SSDs.
This is all taken care of with App Direct mode + DAX. No extra flushes necessary. Things have come a long way over the past couple of years.
@@malventano First off, that was a fun video.
Sorry if I came off like I suggesting it's not a solved problem, that wasn't my intent. My point is that it is solved, exactly like dealing with buffered drives is a solved problem. Same reasons for having buffers, causing the same problems, leading to the same solutions.
I co-wrote a patent for a phase change memory in a DIMM in 2009 so I'm a fan of the concept. Nevertheless, I've long thought pmem advocates do a disservice to the cause by running with talking points that bury the lede. Pmem shines plenty in low latency and small operations, no reason to pick fights you can't easily win. Saying pmem is better because it doesn't buffer data, is just inaccurate and frankly unnecessary. Integration of pmem tech into SW stacks is hard, the ecosystem has made it more accessible, but the point is it can be well worth the effort.
I don't quite get the interest in the Samsung MZ-1LV9600 ... kinda old news (on a 2022 video) ..?
But these guys know this subject better than I (by far) so I assume I'm who's missing something... but what..?
Level1 mini-CES, I'll take that
and you'll take my _engagement_
Around 2 years ago Linus dropped a 10k CPU. Today I have a system with two Intel Xeon platinum p-8124 cpu's. In four year if I'm not playing minecraft Java edition in Linux mint on equivalent to this system I'll be very disappointed in myself.
I see Allyn, I know its going to be insane and cost a ton! Lol great video!
I wanted to hear the full power of deathstar on the table, you should have spun it up.
Wow a whole 88 lanes? And you only need two processors? *laughs in Epyc*
I think Allyn can single-handedly go against all of the bad press Intel has gotten over the last decade
Lovely people with lovely hardware.
Yes, bring Wendell the CES
Holy schnikeys that's a pricey server. Enterprise hardware costs always blow my mind. You could have so many Chrome tabs with all that RAM.
Does this break the Von Neumann architecture requiring different compilers as storage is really memory and data? Couldn't your data and instruction memory be the same thing? Just curious likely I am not understanding it.
Lots of changes need to happen to get to what you’re describing. We might be there today if PMEM tech came about say 20 years ago.
finaly proper Anime storage
Excuse me, can I shout out to the Editor, for handy little on screen definitions of the causal jargon terms please? Thank you
@3:00 Only Oracle? HANA!
Still a pcper fan,
It's a half drunk ramble about tech every week over there
I'm checking them out again now. Didn't like Jim at all, but he seems to have left, so maybe I can binge and re-sub :-D
Great video guys :)
Come on.. Turn it on and show some benchmarks and real life performance 🥺🥺
Why do you need this kind of a server? It is useless for OLAP because 4TB is almost nothing in terms of disk space. It might be a valid platform choice for a search engine, in-memory queue or an OLTP workload, but why do you need these servers unless they are not so expensive? It feels like you still need checksums because even ECC memory fails when you have 100k machines.
Core baby
YAY!!
Can you use these second-gen Optane DIMMs on AMD platforms? 🙃
I want this in my homelab!!!!!!
250K you say? I'll have three.
Love it!
Who are you calling a retimer?
so.. I should continue using my 16 core X299 system then. :)
Remind me of the riovista from long ago
Fuck, that shit neat dude.
I looked I couldn't find any, can some one smarter then me make a sata power loss protection adapter for drives without the caps internally?
Allyn is worth Gold Good sir.
They are 3 bin´s higher, ie 300 mhz. I´m like what you count 100 hz as one bin lol.
That's a lot of cake
Is allyn gonna put this into his Tesla?
Technically it was (I had to get it there somehow :) )
@@malventano Why didn't you "travel by telecom rack"? Is PMem too delicate or were you afraid of ending up in LMG Studios?
Referencing this BTW (When Wendell helped Gamers Nexus build a new NAS): ua-cam.com/video/Hix0l8cFaMw/v-deo.html (this is the punchline of a 2min joke)
i wanted to hear the fans scream
Memhive seems like a good day 1 use of pmem
Server? Dont even know what i would use one for. Do i need it? YES :D
So, i want to download the interweebs. This is the tool i need? And yes, i'm dumb.
woop
Luckily LTT is not there to drop it. :)
*claps*
I like when you have allyn on :)
Its feels almost like being visited by someone from the future lol
pleb tier developer here: so I basically I don´t need Redis with p mem?
Can this thing run a minecraft server?
This is like a Netflix edge server.
But wait -- really .. ? In "10 years this won't be a couple grand on eBay" ..?
I'm skeptical of that. Highly in fact. I think tha'ts exactly what it'll be by then. (at most)
Sadly, I think it's time to pull the plug on this channel. Between the recent video discussing the pope and pets to this one talking about hardware fantasies at a time when I cannot get mainstream memory, I don't see the relevance of the channel anymore. I'm not even sure who this channel is for anymore and I've been in IT since the 70s. Best of luck in the future.
Wendel do you work at Intel too? You always joke that you're janitorial but you are knowledgeable, I wonder what you do professionally.