Preparing for the .NET 10 GC
maoni0.medium.com
maoni0.medium.com
https://learn.microsoft.com/en-us/dotnet/standard/garbage-co...
Then I realized, "oh, it's hosted on Medium." (I generally find Medium posts to be very low quality.) In this case, the author implies that they are on the .Net team, so I'm continuing to read.
(At least I hope the author actually is on the .Net team and isn't blowing hot air, because it's a Medium post and not something from an official MS blog.)
Therefore she's probably the person with the most knowledge about the .net GC but maybe not the best writer (I haven't read the article yet).
Actually, thinking about this more, this is super helpful for Stage VMs that are shared among multiple projects. It can help pack more instances in those VMs.
I have a hard time finding this approach compelling. The amount of additional GC required in their example seems extreme to me.
Posting from a Microsoft blog would to some extent fix this, to the OP’s point.
(I know - who cares. But first impressions are what they are)
I'm just a sometime mid-level .net dev, and I immediately recognized her name.
Here is an interview with her,
https://www.youtube.com/watch?v=ujkSnko0JNQ
Having said this, I agree with you, the Aspire/MAUI architects do the same, I really don't get why we have to search for this kind of blog posts on other platforms instead of DevBlogs.
Many times I have found an interesting article only for it to be just not accesible.
Maybe that is why people stoped using it.
As if they would behave any better, assuming they don't go bankrupt.
I am telling you why they don't use Microsoft blogs, no why they choosed Medium.
Maoni Stephens is the lead developer on the .NET garbage collector. An "About" entry would probably help, but she has a lot of name recognition in the .NET community and in the article it's clear from the first sentence that she's talking from the perspective of owning the GC.
GC: Garbage Collector
DATAS: Dynamic adaptation to application sizes
UOH: User Old Heap. I can't find an explanation for what this is.
LOH: Large Object Heap. This is where allocations over a size threshold go in .NET.
POH: Pinned Object Heap. Pinning is used to stop an object in the GC's memory from being moved around by the GC (for compaction).
ASP.net: Active Server Pages for .NET. This is a framework for building web applications using .NET, a successor to the classic ASP which was built on COM and scripting languages like JScript/VBScript.
Workstation / Server GC: .NET has two major GC modes which have different configurations for things like having per-cpu-core segregated heaps, doing background or foreground GCs, etc. This is designed to optimize for different workloads, like running a webserver vs a graphical application.
Ephemeral GC / Ephemeral generation: To quote the docs:
> For small objects the heap is divided into 3 generations: gen0, gen1 and gen2. For large objects there's one generation – gen3. Gen0 and gen1 are referred to as ephemeral (objects lasting for a short time) generations.
Essentially, generation 0 or gen0 is where brand new objects live. If the GC sees that gen0 objects have survived when it does a collection, it promotes them to gen1, and then they will eventually get promoted to gen2. Most temporary objects live and die in gen0.
Pause time: Most garbage collectors will need to pause the whole application in order to run, though they may not need the application to stay paused the whole time they are working. So pause time and % pause time track how much time the application spends paused for the GC to do its job; ideally these values are low.
BCD: Quoting the post:
> 1) introduced a concept of “Budget Computed via DATAS (BCD)” which is calculated based on the application size and gives us an upper bound of the gen0 budget for that size, which can approximate the generation size for gen0
Essentially, this is an estimate of how much space the ephemeral generation (temporary objects plus some extras) is using.
TCP: Quoting the post again:
> 2) within this upper bound, we can further reduce memory if we can still maintain reasonable performance. And we define this “reasonable performance” with a target Throughput Cost Percentage (TCP). This takes into consideration both GC pauses and how much allocating threads have to wait.
First, I explain that garbage collected applications don't release memory immediately. Then I get sucked into a wild goose chase looking for a memory leak that doesn't exist. Finally, I point out that the behavior they see is normal, usually to some grumbling.
From what I can tell, DATAS basically makes a .Net application have a normal memory footprint. Otherwise, .Net is quite a pig when it comes to memory. https://github.com/GWBasic/soft_matrix, implemented in Rust, generally has very low memory consumption. An earlier version that I wrote in C# would consume gigabytes of memory (and often run out of memory when run on Mono with the Bohem garbage collector.)
---
> If startup perf is critical, DATAS is not for you
This is one of my big frustrations with .net, (although I tend to look at how dependency injection is implemented as a bigger culprit.)
It does make me wonder: How practical is it to just use traditional reference counting and then periodically do a mark-and-sweep? I know it's a very different approach than .net was designed for. (Because they deliberately decided that dereferencing an object should have no computational cost.) It's more of a rhetorical question.
This is what CPython does. The trade off is solidly worse allocator performance, however. You also have the reference counting overhead, which is not trivial unless it is deferred.
There is always a connection between the allocator and collector. If you use a compacting collector (which I assumed .NET does), you get bump pointer allocation, which is very fast. However, if you use a non-compacting collector (mark-and-sweep is non-compacting), you would then fallback to a normal free list allocator (aka as "malloc") which has solidly higher overhead. You can see the impact of this (and reference counting) in any benchmark that builds a tree (and therefore is highly contended on allocation). This is also why languages that use free list allocation often have some sort of "arena" library, so they can have high speed bump pointer allocation in hot spots (and then free all that memory at once later on).
BTW, reference counting, malloc/free performance also impact Rust, but given Rust's heavy reliance on the stack it often doesn't impact performance much (aka just doing less allocations). For allocation heavy code, many of us use MiMalloc one of the better malloc/free implementations.
FWIW: When I look at Azure costs, RAM tends to cost more than CPU. So the tradeoffs of using a "slower" memory manager might be justified.
The problem is in languages like C#/Java almost everything is an allocation, so I don't really think reference counting would work well there. I suspect this is the reason PyPy doesn't use reference counting, it is a big slowdown for CPython. Reference counting really only works well in languages with low allocations. Go mostly gets away with a non-compacting mark-sweep collector because it has low level control that allows many things to sit on the stack (like Rust/C/C++, etc.).
IE, the critical difference is that reference counting frees memory immediately; albeit at a higher CPU cost and needing to still perform a mark-and-sweep to clear out cyclic references.
yes, this is an easily overlooked point: Using memory when it going free is by design. It is often better to use use up cheap, unused memory instead of expensive CPU doing a GC. When memory is plentiful as it often is, then it is faster to just not run a GC yet.
You're not in trouble unless you run short of memory, and a necessary GC does not free up enough. Then only can you call it an issue.
There are some other things that would probably be improvements coming along with refcounting though - you might be able to get rid of GC write barriers.
The most common one I see is LOH (Large Object Heap) fragmentation. When objects are promoted to the LOH the runtime doesnt bother with moving them around anymore. There is a way to explicitly compact the LOH but it can be a non-starter for a lot of applications.
https://learn.microsoft.com/en-us/dotnet/api/system.runtime....
I've once exposed this as a button that a customer's IT department could click whenever they received an alert on memory utilization. The actual solution would have been to refactor the entire product to not pass gigantic blobs around all the time, but that wasn't in the cards for us.
In Server environments. DATAS is an upgrade to garbage collection in "Server mode". Server GC assumed it could be the only thing running on a machine and could use as much memory as it wanted and so would just easily over-allocate memory much more than what it immediately needed. (As the article points out, it would start at a large fixed amount of memory times the number of CPU cores.)
(As opposed to "Workstation GC" which has always tried to minimize memory consumption because it assumes it is running as only one of many apps on an end user system.)
> (and often run out of memory when run on Mono with the Bohem garbage collector.)
Not exactly a fair comparison between .NET's actual GC and Mono's old simpler GC before the merger. (Today's .NET shares the same GC on Windows and Linux [and macOS].)
> This is one of my big frustrations with .net, (although I tend to look at how dependency injection is implemented as a bigger culprit.)
Startup times have gotten a lot better in recent versions of .NET, AOT compiling has much improved (especially compared to the ancient ngen for anyone old enough to remember needing to use that for startup optimization), and while I agree .NET has seen a lot of terrible DI implementations the out-of-the-box one in Microsoft.Extensions does a lot of things right now, including avoiding a lot of Reflection in standard usage which was the big thing slowing down older DI systems. (I've seen people add Reflection based "helpers" back on top of the Microsoft.Extensions DI, but at that point that is a user problem more than a DI problem.)
> It does make me wonder: How practical is it to just use traditional reference counting and then periodically do a mark-and-sweep?
Technically the "mark" of "mark-and-sweep" can be implemented as traditional reference counting (and some of the earliest "mark-and-sweep" implementations did just that). It still only solves half the problem, though. Also, the optimizations made by modern "mark" systems come from that you don't need detailed counts, you just need tools equivalent to Bloom filters (what's the probability this is referenced at least once) and those can be much faster/more efficient to compute and use a lot less memory space than reference counters while doing that.
If your concern is total memory consumption, traditional reference counting uses more space (if only just to store counts), and by itself doesn't solve fragmentation (the "sweep" part of "mark-and-sweep"). From a practical standpoint, combining "traditional reference counting" and a "mark-and-sweep" sounds to me like asking for a less efficient "mark-twice-and-sweep" algorithm.
The important point:
> IE, the critical difference is that reference counting frees memory immediately; albeit at a higher CPU cost and needing to still perform a mark-and-sweep to clear out cyclic references.
Regarding:
> If your concern is total memory consumption, traditional reference counting uses more space (if only just to store counts)
But it also frees memory immediately, meaning that many processes will appear to use less memory (unless fragmentation is an issue.)
Don't forget that GC often adds memory overhead too: IE, mark and sweep sets a generation counter in each object that it can reach, and then objects that weren't updated are reclaimed.
I think where we disagree is I that I of course do assume fragmentation is an issue, and also maybe what "immediately" means in this case. The type of total memory consumption that matters when you look in say Task Manager is when entire pages of memory are returned to the OS, not when individual objects are marked unused/free. In practical concerns, fragmentation will always delay entire pages returning to the OS. Reference counted languages build a lot of tricks to avoid fragmentation sure, but then if you are also trying to use a "mark-and-sweep" heap you lost most of those optimizations in part because you are then already assuming fragmentation is a problem to solve.
> Don't forget that GC often adds memory overhead too: IE, mark and sweep sets a generation counter in each object that it can reach
I did mention it, but also that GCs have advanced from "include a generation counter in each object" to things like generation bitmaps where that data is stored outside of the objects themselves and then from there further optimized into even more "compressed" forms ala Bloom filters (they maybe don't track every object, but every cluster of objects, or just objects crossing generation boundaries, or they use hash buckets and probability analysis, and many of these structures don't need to be permanent but are transient only during specific types of garbage collections; there has been a lot of work in the space and many decades of efficiencies studied and built). It's still overhead, but it is now a very different class of overhead from reference counts.
Yes, I'm very aware of that.
Remember, my question about reference counting's practicality in C# is more of a rhetorical question to encourage discussion.
GC? -- Maybe "Garbage Collection", i.e., have some memory (mainly computer main memory) allocated, don't need it (just now or forever), and want to release it, i.e., no longer have it allocated for its original purpose. By releasing can make it available for other purposes, software threads, programs, virtual machines, etc.
DATAS? -- Not a spelling error or about any usual meaning for data and instead is as in
https://learn.microsoft.com/en-us/dotnet/standard/garbage-co...
for "Dynamic adaptation to application sizes"
So, we're trying to take actions over time in response to some inputs that are in some respects unpredictable.
Okay, what is the objective, i.e., the reason, what we hope to gain, or why bother?
And for the part that is somewhat unpredictable over time, that's one or more stochastic processes (or one multidimensional stochastic process?).
So, in broad terms, we are interested in stochastic optimal control. "Dynamic adaptation", is close and also close to one method, dynamic programming -- in an earlier thread at Hacker News, gave a list of references. Confession, wrote my applied math Ph.D. dissertation in that subject.
Hmm, how to proceed??? Maybe, (A) Know more about the context, e.g., what the computer is doing, what's to be minimized or maximized. (B) Collect some data on the way to knowing more about the stochastic processes involved.
For me, how to get paid? If tried to make a living from applied stochastic optimal control, would have died from starvation. Got the Ph.D. JUST to be better prepared as an employee for such problems and had to learn that NO one, not even one in the galaxy, cares as much as one photon of ~1 Hz light.
So, am starting a business heavily in computing and applied math. The code from Microsoft tools is all in .NET, ASP.NET, ADO.NET, etc. Code runs fine. The .NET software, via the VB.NET syntactic sugar, is GREAT for writing the code.
So, MUST keep up on Microsoft tools, and here just did that. Since .NET 10 is changing some versions of Windows, my reaction is (i) add a lot of main memory until GC is nearly irrelevant, (ii) in general, wait a few years to give Microsoft time to fix problems, i.e., usually be a few years behind the latest versions, i.e., to "Prepare for .NET 10", first wait a few years.
Experience: At one time, saw some server farms big on reliability. One site had two of everything, one for the real work and another to test the latest for bugs before being used for real work. Another had their own electrical power, Diesel generators ~30 feet high, a second site duplicating everything, ~400 miles away, with every site with lots of redundancy. In such contexts, working hard and taking risks trying to save money on main memory seem unwise.