We Replaced an SSD with Storage Class Memory

https://engineering.mongodb.com/post/we-replaced-an-ssd-with-storage-class-memory-here-is-what-we-learned

by ssvss • 6 years ago

125 68 6 years ago

engineering.mongodb.com

Jesus Christ, this is insane. Almost a Terabyte of 12.6 Gbps reads? I have a bunch of geospatial entity resolution workloads that I could absolutely smash with this. For way cheaper than the fat mem instances.

lrem3• 6 years ago

Back in the day when I interviewed for Google I had this beautiful question. The interviewer fished for a basic distributed key-value store. I just kept coming up with single machine solution to his numbers. "No, I really can have that storage+bandwidth, here's the part number."

I'm still wondering if that interview costed me a level.

lmilcin6• 6 years ago

Being interviewer myself I can say interviewers tend to be stubborn. They have an excellent question (at least they think they have) that they have spend a lot of time thinking through and discussing with other candidates. Your answer causes him/her to miss the opportunity to have a real discussion and understand your knowledge of the topic.

It is better to play along and only mention the real solution at the end, to finish on a high note.

ur-whale• 6 years ago

If the interviewer was gauging your political skills/ adaptability, it doesn't look like you fared too well indeed.

asdfasgasdgasdg2• 6 years ago

Even if you can do a particular problem on a single machine, sometimes that isn't the right call. In a work scheduled cluster environment, a task that wants the entire machine may have trouble getting a slot unless it has the priority to preempt everything else going on on that machine. We call such VMs "picky" and they don't get scheduling guarantees.

dominotw• 6 years ago

solution to his numbers.

Why did he have concrete numbers. Couldn't he simply have increased those numbers when you proposed single machine solution.

twic• 6 years ago

An interviewer once gave me a problem about scraping some information out of a log file. I wrote a short shell pipeline. The interviewer asked for a more complicated analysis. I wrote a longer pipeline. The interviewer added more requirements, with aggregation, state, etc in order to push me to write an actual program. I wrote an even longer pipeline.

By this point we had both realised that this was a battle of wits: could he come up with a problem that i couldn't solve with a pipeline?

At the end of the interview, i had a pipeline that took up most of a piece of A4 paper to write out. I had won the battle, and was offered the job.

Of course, i would not advise you to actually write a pipeline like that in production, but it's a fun exercise.

Anyway, the moral of this story is that if the interviewer wants you to solve a problem a certain way, and you can solve it in a simpler way, then a good interviewer will mark you up, not down. Perhaps at Google they didn't; they don't really seem like a company that has it together.

sukilot6• 6 years ago

Your interviewer may have lacked skill in guiding the process, but you could have taken the question in the spirit it was given and answered "now if the usage grows 10x as much as hardware advances" or "if we need to be resilient to hardware failures" or "if we need to roll out a logic upgrade slowly" and continued to solve the problem.

I'm still wondering if that interview costed me a level.

Unlikely. Candidate level is decided before the interview.

jiggawatts2• 6 years ago

GBps, not Gbps! They were getting 12.6 gigabytes per second, which would be 100 gigabits per second almost exactly.

However, even their 6-DIMM test produces only 300 Gbps, which is insufficient to saturate a modern 400 GbE network adapter for either reads or writes.

This would be most relevant on a "single master" system storing some sort of simple data where consistency requirements means that the "writer" cannot be distributed. In a situation like this, the NIC and the storage bandwidth are the ultimate limits.

In general, Intel SSDs and Intel Optane have poor but consistent bandwidth, and consistently low latency. Coupled with the high price and small capacity, they have their niche, but they're not a clear winner in any category.

As a reference point for how crazy high bandwidths are these days, NVIDIA sells a turnkey solution with 200 GB/sec network bandwidth (1.6 Tbps!): https://www.nvidia.com/en-au/data-center/dgx-a100/

lmilcin6• 6 years ago

Where this is useful is when the application needs to process much more data than it then needs to transmit over the network.

For a database system like mongodb this could be perfect depending on workload.

Jabbles• 6 years ago

Does anyone have good numbers for a) RAM, and b) the cost of each of these solutions?

(RAM is obviously volatile.)

Pelic4n2• 6 years ago

So you need to do THAT to get decent perfs with MongoDB. Good to know!

jiggawatts2• 6 years ago

Unless I'm missing something, their benchmark graphs at the bottom of the report show that there is no significant benefit to using SCM with MongoDB! Their internal overheads must be high enough to swamp the few microseconds gained from the faster storage.

Pelic4n2• 6 years ago

That was a joke. I've been using MongoDB in production for years and I'm salty. :)

fortran77• 6 years ago

That was my first thought, too. You need these tricks to make MongoDB "WebScale"

http://www.mongodb-is-web-scale.com/

(I can get better results for Key-Value storage by using an SQL Database--Postgres or MariaDB--for key/values over MongoDB. And if you know SQL well, you can get even better results using a real relational database and optimal queries than pulling out keys/values and ad-hocking your relations in some Javascript code or whatever these web kids are doing.)

jd_mongodb• 6 years ago

You don't pick a database like MongoDB purely for its key/value query performance. SQL will do really well until it does really badly, for example in cases where you exceed the limitations of the box/VM it is running on.

lichtenberger4• 6 years ago

You can efficiently read 256 Byte granular data (4 cache lines) with Optane Memory (due to checksums). I think it makes much more sense to read/write fine granular changes for instance at least align pages to 64 or 256 Bytes instead of 4kb pages, where you often times first of all write too much data and secondly you pollute the caches with probably unnecessary data. There's a paper about how to add cache line aligned mini-pages (16 cache-lines): https://db.in.tum.de/~leis/papers/nvm.pdf

ayende• 6 years ago

Writing on a page (4KB / 8KB) boundary is almost always a good idea. Because there is an overhead per page that you have to account for. It can be in the range of 16 - 64 bytes, so having a 256 mini page is probably a bad idea.

Most of the _data_ is also not going to fit in 256 bytes anyway.

lichtenberger4• 6 years ago

It might add some overhead, but I guess it depends on the page implementation, but 16 bytes seems to be the minimum (and Optane Memory might , I agree. That said if someone changes only one record the best thing is to write in the smallest granularity possible on the storage medium.

So, it might well be that someone is only interested in only one or a few records. Why then fetch and cache a whole 4Kb page if latency is good in both cases (4kb and 256 bytes)? On the other hand I agree that you should probably cache more data from a hot page.

lichtenberger4• 6 years ago

I think along those lines, reading and writing fine granular data, especially when storing the full history of the data is the way to go (in addition to path copying and a log-structured storage). Furthermore pointer swizzling might be needed to get almost in-memory database performance.

I'm trying to address this with my versioned Open Source DBMS project, where only page-fragments are stored and bunch of them is read in-parallel to reconstruct a full page in-memory. Adding mini-page ps plus a simple cache at some point with hot data from several page fragments at least is orthogonal:

https://github.com/sirixdb/sirix | https://sirix.io

wtallis• 6 years ago

FYI, Intel Optane Memory is a wildly different product from the one under discussion, which is Optane DC Persistent Memory or Optane DCPMM. The former is a low-capacity consumer SSD caching software solution, and the latter is persistent memory modules in a DIMM form factor.

lichtenberger4• 6 years ago

Yes, I mean Optance DC Persistent Memory. It's hard because they seems to market three different product categories under almost the same name (Optane Memory, Optane DC Persistent Memory and SSDs).

cm2187• 6 years ago

Stupid question. What are the use cases for such massively fast write speeds?

If you are storing data to disk at that speed, you fill even the biggest optane drives in a couple of minutes. So it would be an application where you need to overwrite a huge amount of data over and over again.

stonewhite• 6 years ago

That also necessitates a matching network bandwidth as well.

It reminds me of the Bugatti Veyron situation, its tires lasting 15 minutes at top speed yet it runs out of fuel in 12.

falcolas2• 6 years ago

I'd rather run out of fuel and drift to a stop, instead of having a blowout and crash while driving at top speed. I'm not sure where the problem lies.

dr_zoidberg• 6 years ago

Digital forensics is a situation where you need extreme IO but you aren't really writing all that much -- it's mostly searches over the forensic copy of computers you are analizing.

As for write speed, you don't really need it unless you're doing file/data recovery, but even then you can just make indexes to the disk image.

AFAIK, very few (if none) are using SCM in their forensic expert workstations, so I can't really tell if you'd saturate the storage capacity or the bandwidth. But in digital forensics, NVMe's have been a must for a while, so this would definitely help.

lmilcin6• 6 years ago

A lot of problems don't scale that easily. You may want to prolong the life of the solution by just giving it faster storage. Think in terms of giving faster storage for Oracle DB vs migrating your app to some better scaling nosql solution. SCM might seem expensive until you see the bill for the development costs.

Another use is to have local storage. You want most of the communication to be between your CPU and RAM/SCM and networking only to transmit results. In case of mongodb that would be something like very heavy aggregation that needs to look at a lot of data but transfers relatively small amount of results to the client (think counting number of objects meeting an arbitrary set of specifications). MongoDB actually requires storage for writing intermediate results of some types of operations so it is not purely about read speed.

deepstack• 6 years ago

Timeseries db is one thing that comes to mind.

yxhuvud• 6 years ago

You can for example flush a temporary db table to storage while doing computations if it ends up too big to fit in memory. Just because it is fast doesn't mean it need to be sustained throughput.

georgewfraser• 6 years ago

Andy Pavlo talks about this in his class at CMU. You shouldn’t expect to get better performance by running a disk-optimized storage engine on memory, because you’re still paying all the overhead of locks and pages to work around the latency of disk, even though that latency no longer exists. Instead, you have to build a new, simpler storage engine that skips all the bookkeeping of a disk-oriented storage engine.

https://youtu.be/a70jRWLjQFk

mackle_hair• 6 years ago

i believe there are other issues with SCM arent there, for instance if you flip the same bit over and over, you'll wear out the media.. so just like with SSD's there has to be some wear leveling algorithms, but because the media has gotten so fast, the wear leveling would theoretically become a larger % of the bottleneck of using the drive.

but i totally agree, rethinking how data is stored is going to be key to the adoption of these types of new media. there's a company called vast data that has built out a full storage solution utilizing the unique properties of SCM, very cool https://vastdata.com/

pocket_cheese• 6 years ago

For anyone interested in how a database works underneath the hood, I could not recommend a better MOOC than Andy Pavlo's database courses. His intro to database course is so freaking good.

We Replaced an SSD with Storage Class Memory

Related Stories