I think they’re playing it safe with their points of comparison. DDR5 supports 64,000 MB/s per channel and 4x NVMe PCIe5 SSDs support 10,000 MB/s. Depending on how many memory channels and what RAID you use, I think 10x improvement over high performance storage is unimpeachable. Memory latency being better than SSD latency will really benefit memory depending on workload, but I don’t think you can just throw one number out there to represent that.
Now if you’re comparing to spinning rust, memory is definitely going to blow it away, but commodity hardware isn’t running tens or hundreds of TBs of memory. Memory to SSD comparisons seem right.
For example the latest AMD Genoa has 12 memory channels per socket, at dual socket and with enough DIMMs that's a 75x speed advantage even if you compare with RAID0 of high perfomance NVMe.
> even if you compare with RAID0 of high perfomance NVMe.
Do you mean a RAID0 of just two or four NVMe SSDs? It's absolutely ridiculous to count aggregate DRAM bandwidth across two CPU sockets and not do the same for PCIe lanes. A fair comparison is that Genoa has about twice the DRAM bandwidth as it has PCIe bandwidth, though in a fully-loaded database server some portion of the PCIe bandwidth will be used for networking rather than storage.
FWIW I think you have the aggregation backwards, dual socket Genoa would have 24 channels of DRAM but sacrifice some of the PCIe lanes for the interconnect. Your numbers work out right though as single socket is actually about 1:1 in RAM vs PCIe bandwidth so dual socket would still come out roughly 2:1.
I think 10x is a fair rough number though, depending on your access pattern.
Dual-socket Genoa would be 24 channels at DDR5-4800 (38.4 GB/s) for a total of ~921.6 GB/s. Typical PCIe configurations are 64 or 80 lanes per socket for dual-socket, so 128 or 160 lanes total, at PCIe 5.0 speed that's ~504 GB/s for 128 lanes total or ~630 GB/s for 160 lane configurations.
Single-socket Genoa would be 12 channels of DRAM (~460.8 GB/s) and 128 lanes of PCIe 5.0 (~504 GB/s), but none of the previous comments were specifically about single-socket Genoa and I wasn't going to silently switch from considering dual-socket in one sentence to single-socket in the next sentence.
I guess when you compare pure performance of hardware, 10x is very optimistic from the perspective of disk. Probably the author based this number on some some specific application/database context / bias in the measurements. But yea, pure hardware difference might be hugely different compare to that number.
Based on this resource https://gist.github.com/hellerbarde/2843375 1MB sequential read from SSD just 4x slower than the same read from main memory.
For random reads main memory faster of course.
I believe author meant some average value.
Recent database engine designs tend to be bandwidth bound. The difference in bandwidth between memory and modern storage hardware is much smaller than you might expect. Really taking advantage of that storage bandwidth is more difficult than memory, as it requires some top notch scheduler design.
> here’s a chart comparing the throughputs of typical memory, I/O and networking technologies used in servers in 2020 against those technologies in 2023
> Everything got faster, but the relative ratios also completely flipped
> memory located remotely across a network link can now be accessed with no penalty in throughput
Yeah, I think this is a brilliant thing in part because it makes disaggregation of hardware infrastructure plausible for high-performance, high-scale data-intensive workloads. It also turns some deeply embedded assumptions about software design and architecture for things like databases completely on their head. This shift in hardware balance is probably the biggest change to fundamental assumptions in database architecture since SSDs started to become a thing, and most people haven't internalized the implications yet.
It is an fun time to be working in high-scale data systems.
Yes we used this number based on some of our workloads and it is definitely lower bound... As I am thinking now we should have moved this lower bound a bit up yes
Not 100% what the article is about, just a short story.
In one of my old jobs we had megabytes of infrequently accessed static key-value data. If we simply loaded it into a const (i.e. a hash table), it would blow up the RAM so much, that we would need to upgrade to bigger VPSes. If we put it in database, it would make it annoying to keep these tables up to date, track changes in them.
I figured this was one of those in-between use cases, where the best solution is to have zero-RAM lookups from SSD. In my case, I wrote a little ruby library[1] that arranges data in equal cells in a file, and performs binary searches via `pread`. This was perfect for us, because we kept data in our repos, sacrificed no RAM at runtime, SSD lookups were fast enough, and we didn't have to support a more elaborate db.
Good paper and overall the right opinion but not very relevant to a relatively simple system as OP described since they weren't dealing with safety, transactions, etc.
I'm no systems programmer, but I remember trying to research mmap approach. It's been 3 years, so I'm not sure what stopped me, but something didn't feel right. Perhaps it was the lack of control over how memory is used. I could clearly see how not use it, and didn't want any fluctuations.
Edit: oh and I think I did come across some article like the one linked in the neighbor comment. It's starting to come back.
Sure -- until your system crashes and data is left in an inconsistent state because some of it was written to non-volatile storage and some of it wasn't and your OS had no concept of transactional consistency because mmap is a leaky abstraction.
[UPDATE] I missed a crucial part of the problem setup, which is that the data is read-only. That obviously moots my objection.
Basically, your solution ended up residing on-disk, but the data was fragmented in such a way that efficient lookup was possible? Did this copy from kernel space to userspace, also?
Are you saying that the file would just entirely be loaded into a page cache eventually? I imagine, even if true, it still wouldn't result in OOM killing the server/worker daemons on account of this data?
Depending on the file size and available memory, yes. In the event memory needs to be reclaimed Linux’s memory management system will free from the page cache first if processes need more anonymous memory.
The "log(n) * number of searches" page lookups, that the kernel could clean up at any time after a search, instead of all n items that it would have had to load and keep in memory?
They've misunderstood what an in-memory database is though. The distinction only makes sense in a historical perspective, with on-disk formats and associated algorithms that worked well for execution from spinning iron. In fact neo4j is also an in-memory database, since it will slow to a complete crawl on a HDD. The fact that it supports swapping pages in and out of disk can more be likened to virtual memory, rather than any specific support for executing queries efficiently from disk.
Hey I am Andi, the author of blog post and wanted to jump in... First neo4j is not an in-memory db since its main storage is on disk. It does come with cache mechanisms using which data eventually ends in main memory but that is completely opposite approach of what we wanted to show in this blog post and what Memgraph is capable of doing. Memgraph is an explicite in-memory storage (or was since now it has capability to also use disk) because everything is computed in memory, no caching needed no buffer pool, for in-memory storage the only disk-related thing is securing the durability but that is different topic
Most databases have their main storage on disk. That is how they define the D in ACID, durable. That is not the salient point regarding what an in-memory database is.
I appreciate that your main point is with regards to memgraph internals and it was an interesting read.
Sure, querying that data is also important. Which means the question of “do you optimise for in-memory queries, or on disk queries” is important.
Simply saying you want the data to be durable isn’t that interesting or hard to achieve, there’s plenty of ways of achieving durable storage. The hard part is doing durable storage while also solving problems like query speed, and concurrency control.
It depends on what you mean by durable. Claiming that persistence to disk is trivial when compared to returned values is a stretch IMHO. Modern DBs still have bugs where data returned is different than permanently persisted leading to inconsistency errors.
And that is ignoring performance queries that don't require any guarantees.
Unless you’re writing directly to a flat file, and basically storing a direct representation of your data on disk, as it exists in your application. Then your DB is doing something to improve the performance of your queries, whether that indexes, or other clever on-disk structures, it exists only to improve query performance.
Databases don’t have to use a file-system - look at old Big Iron systems where the database is the file-system.
In the context of Linux/Unix/etc, the file-system is just another API for the OS - consider /dev/null or /proc - those are certainly in the file-system but they aren’t tied to persistent storage.
No, I'm saying that persisting data is mainly what a filesystem is about unless you go into some pedantic nonsense trying to tear apart a general statement.
This is like someone saying memory is about storing data while the computer is on and someone else trying to start an argument by saying that sometimes memory addresses are used to communicate with video cards.
53 comments
[ 3.9 ms ] story [ 121 ms ] threadDid they mean 10 thousand times? Or is the in-memory version that inefficient?
Now if you’re comparing to spinning rust, memory is definitely going to blow it away, but commodity hardware isn’t running tens or hundreds of TBs of memory. Memory to SSD comparisons seem right.
Do you mean a RAID0 of just two or four NVMe SSDs? It's absolutely ridiculous to count aggregate DRAM bandwidth across two CPU sockets and not do the same for PCIe lanes. A fair comparison is that Genoa has about twice the DRAM bandwidth as it has PCIe bandwidth, though in a fully-loaded database server some portion of the PCIe bandwidth will be used for networking rather than storage.
I think 10x is a fair rough number though, depending on your access pattern.
Single-socket Genoa would be 12 channels of DRAM (~460.8 GB/s) and 128 lanes of PCIe 5.0 (~504 GB/s), but none of the previous comments were specifically about single-socket Genoa and I wasn't going to silently switch from considering dual-socket in one sentence to single-socket in the next sentence.
> here’s a chart comparing the throughputs of typical memory, I/O and networking technologies used in servers in 2020 against those technologies in 2023
> Everything got faster, but the relative ratios also completely flipped
> memory located remotely across a network link can now be accessed with no penalty in throughput
The graphs demonstrate it very clearly: https://blog.enfabrica.net/the-next-step-in-high-performance...
It is an fun time to be working in high-scale data systems.
This should have been a revolution in DB design, IMO, but it kind of hasn't been.
In one of my old jobs we had megabytes of infrequently accessed static key-value data. If we simply loaded it into a const (i.e. a hash table), it would blow up the RAM so much, that we would need to upgrade to bigger VPSes. If we put it in database, it would make it annoying to keep these tables up to date, track changes in them.
I figured this was one of those in-between use cases, where the best solution is to have zero-RAM lookups from SSD. In my case, I wrote a little ruby library[1] that arranges data in equal cells in a file, and performs binary searches via `pread`. This was perfect for us, because we kept data in our repos, sacrificed no RAM at runtime, SSD lookups were fast enough, and we didn't have to support a more elaborate db.
[1]: https://github.com/maxim/wordmap
Edit: oh and I think I did come across some article like the one linked in the neighbor comment. It's starting to come back.
[UPDATE] I missed a crucial part of the problem setup, which is that the data is read-only. That obviously moots my objection.
"static key-value data. ... a const (i.e. a hash table)"
Yeah, there's a build step that sorts and arranges data into "cells", making binary search possible.
Not sure about kernel/user space. I'm just calling `pread` from ruby, so only a few bytes are loaded per lookup.
Those lookups eventually made their way into the kernel's page cache.
Yes, they did.
Gives me great performance (especially paired with Go), and I'm able to deliver 6-10k rps for a few dollars a month.
Most databases have their main storage on disk. That is how they define the D in ACID, durable. That is not the salient point regarding what an in-memory database is.
I appreciate that your main point is with regards to memgraph internals and it was an interesting read.
So either way you turn it both solutions go to disk?
Simply saying you want the data to be durable isn’t that interesting or hard to achieve, there’s plenty of ways of achieving durable storage. The hard part is doing durable storage while also solving problems like query speed, and concurrency control.
And that is ignoring performance queries that don't require any guarantees.
You are thinking of a file system.
In the context of Linux/Unix/etc, the file-system is just another API for the OS - consider /dev/null or /proc - those are certainly in the file-system but they aren’t tied to persistent storage.
I didn't say they did.
consider /dev/null or /proc - those are certainly in the file-system but they aren’t tied to persistent storage.
I'm not sure what point you're trying to make here.
> > The whole point of a DB is to remember the data after power cycle. > You are thinking of a file system.
…as though you’re saying only an FS can persist data.