[worstspotgain]

I briefly looked over your stock exchange code:

- For memory management, consider switching to std::shared_ptr. It won't slow anything down and will put that concern to rest entirely.

- For sockets, there are FOSS libraries that will outperform your code and save you a ton of headaches dealing with caveats and annoyances. For example, your looping through FD_ISSET is slower than e.g. epoll or kqueue.

- For dependencies, C++ is definitely wilder than other languages. Dependencies are even harder to find than they are to manage. There's a lot of prospective library code, some of it hidden in little forgotten folds of the Internet. Finding it is basically a skill unto itself, one that can pay off handsomely.

[zxcvbn4038]

When I did low latency everyone was offloading TCP to dedicated hardware.

They would shut down every single process on the server and bind the trading trading app to the CPUs during trading hours to ensure nothing interrupted.

Electrons travel slower than light so they would rent server space at the exchange so they had direct access to the exchange network and didn't have to transverse miles of cables to send their orders.

They would multicast their traffic and there were separate systems to receive the multicast, log packets, and write orders to to databases. There were redundant trading servers that would monitor the multicast traffic so that if they had to take over they would know all of the open positions and orders.

They did all of their testing against simulators - never against live data or even the exchange test systems. They had a petabyte of exchange data they could play back to verify their code worked and to see if tweaks to the algorithm yielding better or worse trading decisions over time.

A solid understanding of the underlying hardware was required, you would make sure network interfaces were arranged in a way they wouldn't cause contention on the PCI bus. You usually had separate interfaces for market data and orders.

All changes were done after exchange hours once trades had been submitted to the back office. The IT department was responsible for reimbursing traders for any losses caused by IT activity - there were shady traders who would look for IT problems and bank them up so they could blame a bad trade on them at some future time.

[shaklee3]

You don't need to shut down processes on the server. All you have to do is isolate CPU cores and move your workloads onto those cores. That's been a common practice in low latency networking for decades.

[bluGill]

I'm not in HFT, but I wouldn't expect that to be enough.

Not only do you want to isolate cores, you want to isolate any shared cache between cores. You do not want your critical data ejected from the cache because a different core sharing the cache has decided it needs that cache. Which of course starts with knowing exactly what CPU you are using since different ones have different cache layouts.

You also don't want those other cores using up precious main memory or IO bandwidth at the moment you need it.

[worstspotgain]

Just to add to your good points: since there's always a faster cache for your working set to not fit in, you can use memory streaming instructions to reduce cache pollution. Depending on the algorithm, increasing cache hit rates can give ridiculous speed-ups.

[shaklee3]

Correct. I was just pointing out to OP that moving processes is not worthwhile and isolation is how you'd do it

[gohwell]

I’ve worked at a few firms and never heard of an IT budget for f-ups. Sounds like a toxic work environment.

[davidmr]

Same. That sounds like a way to make that relationship between front office and back office as toxic and unproductive as possible.

[hawk_]

Depends on how it's set up. You take a chunk of profits as well if things go well.

[resonious]

It's just business, no? Would you rather trade with a service that's liable for their mistakes or one that isn't?

[rramadass]

Any good books/resources you can recommend to learn about the above architectures/techniques?

[neomantra]

Some years ago I wrote a gist about HFT/HPC systems patterns (versus OPs C++ patterns) applied to dockerized Redis. Might be dated, but touches on core isolation/pinning, numa/cgroups, kernel bypass, with some links to go deeper. Nowadays I do it with Kubernetes and Nomad facilities, but same basic ideas:

https://gist.github.com/neomantra/3c9b89887d19be6fa5708bf401...

[rramadass]

Nice; reminds me of the Redhat Performance Tuning and Real Time Low Latency Optimization guides.

[crabmusket]

A few episodes of Signals and Threads, a podcast from Jane Street, go into parts of it.

[rramadass]

Thank You.

[ra0x3]

A great insightful comment, thank you!

[sneilan1]

I did not know std::shared_ptr would not slow things down. I've learned something new today! :)

Yes, I agree, epoll is a lot better than FD_ISSET.

Maybe I can keep moving with my C++ code but do people still trust C++ projects anymore? My ideal use case is a hobbyist who wants a toy stock exchange to run directly in AWS. I felt that C++ has a lot of bad publicity and if I want anyone to trust/try my code I would have to rebuild it in rust.

[worstspotgain]

Here's how to maximize shared_ptr performance:

- In function signatures, use const references: foo(const std::shared_ptr<bar> &p). This will prevent unnecessary bumps of the refcount.

- If you have an inner loop copying a lot of pointers around, you can dereference the shared_ptr's to raw pointers. This is 100% safe provided that the shared_ptr continues to exist in the meantime. I would consider this an optimization and an edge case, though.

I would say people trust C++ projects at least as much as any other professional language - more so if you prove that you know what you're doing.

[Maxatar]

> In function signatures, use const references: foo(const std::shared_ptr<bar> &p). This will prevent unnecessary bumps of the refcount.

This advice doesn't seem quite right to me, and in my codebases I strictly forbid passing shared_ptr by const reference. If you don't need to share ownership of bar, then you do the following:

    foo(const bar&);

If you do need to share ownership of bar, then you do the following:

    foo(std::shared_ptr<bar>);

Why do we pass by value when sharing ownership? Because it allows for move semantics, so that you give the caller to option to make a copy, which bumps up the reference count, or to entirely avoid any copy whatsoever, which allows transfering ownership without bumping the reference count.

Having said that, shared_ptrs do have their uses but they are very very rare and almost all of our use cases do not expose shared_ptr's in the public API but rather use them as an implementation detail. We use them almost exclusively for things like immutable data structures, or copy-on-write semantics, or as a part of a lock-free data structure.

[spacechild1]

> If you don't need to share ownership of bar, then you do the following: > > foo(const bar&);

Exactly!

> This advice doesn't seem quite right to me, and in my codebases I strictly forbid passing shared_ptr by const reference

There is at least one use case I can think of: the function may copy the shared_ptr, but you want to avoid touching the reference count for the (frequent) case where it doesn't. This is an edge case, though, and personally I almost never do it.

[vitus]

Additionally: if you care about nullability semantics within your function, then you write foo(const bar*) and pass in bar_ptr.get(), and of course check that the value is != nullptr before dereferencing it.

Otherwise, I'm inclined to agree -- don't pass around smart pointers unless you're actually expressing ownership semantics. Atomics aren't free, ref-counting isn't free, but sometimes that genuinely is the correct abstraction for what you want to do.

One more point: shared ownership should not be used as a replacement for carefully considering your ownership model.

(For readers who might not be as familiar with ownership in the context of memory management: ownership is the notion that an object's lifetime is constrained to a given context (e.g. a scope or a different object -- for instance, a web server would typically own its listening sockets and any of its modules), and using that to provide guarantees that an object will be live in subcontexts. Exclusive ownership (often, in the form of unique_ptr) tends to make those guarantees easier to reason about, as shared ownership requires that you consider every live owning context in order to reason about when an object is destroyed. Circular reference? Congrats, you've introduced a memory leak; better break the cycle with weak_ptr.)

[worstspotgain]

Let me preface by noting that I don't necessarily disagree with any part of what you wrote. However, there are design patterns that exceed the guardrails you're thinking of, and those are the patterns that benefit the most from shared_ptr.

Typically, they involve fine- to medium-grained objects, particularly those that have dynamic state (meaning by-value copies are not an option.)

An example might be a FlightAware-like system where each plane has a dynamically-updated position:

   class Plane { ... void UpdatePosition(const Pos &); Pos GetPosition() const; };
   using PlanePtr = std::shared_ptr<Plane>;
   using PlaneVec = std::vector<PlanePtr>;
   class Updater { ... PlaneVec mPlanes; };
   class View { ... PlaneVec mPlanes; };

Updater routinely calls UpdatePosition(), whereas View only calls const methods on Plane such as GetPosition(). There can be a View for, say, Delta flights and one for United. Let's simplify by assuming that planes are in the sky forever and don't get added or removed.

Destructing Updater doesn't affect Views and vice-versa. Everything is automatically thread-safe as long as the Pos accesses inside each Plane are thread-safe.

The key here is that Plane is fine-grained enough and inconsequential enough for lazy ownership to be ideal.

[worstspotgain]

> This is an edge case, though, and personally I almost never do it.

My experience is the opposite. It has to do with the coarseness of the objects involved and the amount of inter-object links. We typically have a vast variety of classes. Many of them have shared_ptr members, resulting in rich graphs.

Many methods capture the shared_ptr parameters by copying them inside other objects. However, many methods just want to call a couple methods on the passed-in object, without capturing it. By standardizing on const shared_ptr &, all calls are alike, and callees can change over time (e.g. from not capturing to capturing.)

[worstspotgain]

foo(const bar&) is ideal if you precisely wish to bar ownership. If (and in many kinds of projects, invariably it's more like when) you later decide to share ownership, or if nullptr is an option, then it's no good.

foo(std::shared_ptr<bar>) is copy-constructed as part of your function call (bumping the refcount) unless copy elision is both available and allowed. It's only ideal if you almost always pass newly instantiated objects.

Pass by const reference is the sweet spot. If you absolutely must minimize the refcount bumps, overload by const reference and by rvalue.

As for shared_ptrs being very rare, uh, no. We use them by the truckload. To each their own!

[CyberDildonics]

foo(const bar&) is ideal if you precisely wish to bar ownership.

What?

invariably it's more like when) you later decide to share ownership,

shared_ptr shouldn't even be necessary for keeping track of single threaded scope based ownership.

As for shared_ptrs being very rare, uh, no. We use them by the truckload. To each their own!

You might want to look into that, you shouldn't need to count references in single threaded scope based ownership. If you need something to last longer, make it's ownership higher scoped.

If something already works it works, but this is not necessary and is avoiding understanding the actual scope of variables.

[worstspotgain]

> What?

The context you're missing is in your post's GP. That poster holds a std::shared_ptr<bar> (for whatever perfectly valid reason) and wishes to pass it to foo(). However, he declares it as foo(const bar &) because the callee does not need to share in the ownership of the shared_ptr. That means it gets called as foo(*p.get()).

> scope based ownership

That's the incorrect assumption that you came to. Obviously if bar is only used for stack-based scope variables, no shared_ptr is needed.

[quotemstr]

> Why do we pass by value when sharing ownership? Because it allows for move semantics, so that you give the caller to option to make a copy, which bumps up the reference count, or to entirely avoid any copy whatsoever, which allows transfering ownership without bumping the reference count.

What if the callee sometimes wants to get a reference count and sometimes doesn't? In the latter case, your proposed signature forces an unnecessary pair of atomic reference count operations. But if you use

    foo(bar const&)

instead, then foo can't acquire a reference even when it wants to.

You could stick std::enable_shared_from_this` under `bar`. But `std::enable_shared_from_this` adds a machine word of memory, so you might not want to do that.

If you pass

    foo(shared_ptr<bar> const&)

you incur an extra pointer chase in the callee. Sure, you could write

    foo(bar const&, shared_ptr<bar> const&)

but then you burn an extra argument register. You can't win, can you?

You can win actually. Just use https://www.boost.org/doc/libs/1_85_0/libs/smart_ptr/doc/htm... or your favorite intrusive reference-counted smart pointer, not `std::shared_ptr`. If you do, you get the same capabilities that `std::enable_shared_from_this` grants but without any of the downsides.

[worstspotgain]

> If you pass

   > foo(shared_ptr<bar> const&)

> you incur an extra pointer chase in the callee.

Actually this is usually not the case (assuming of course that caller is holding the original pointer in a shared_ptr<bar> which is the use case we were discussing.)

That shared_ptr<bar> instance is held either on the stack (with address FP + offset or SP + offset) or inside another object (typically 'this' + offset.) To call foo(const shared_ptr<bar> &), the compiler adds the base pointer and offset together, then passes the result of that addition - without dereferencing it.

So as it turns out, you may actually have one fewer pointer chase in the const shared_ptr<bar> & case. For example, if foo() is a virtual method and a specific implementation happens to ignore the parameter, neither the caller nor the callee ever dereference the pointer.

The one exception is if you've already resolved the underlying bar& for an unrelated reason in the caller.

I do agree that intrusive_ptr is nice (and we actually have one codebase that uses something very similar.) However shared_ptr has become the standard idiom, and boost can be a hard sell engineering-wise.

[quotemstr]

> That shared_ptr<bar> instance is held either on the stack (with address FP + offset or SP + offset) or inside another object (typically 'this' + offset.) To call foo(const shared_ptr<bar> &), the compiler adds the base pointer and offset together, then passes the result of that addition - without dereferencing it.

You're overthinking it. Think in cache lines. No matter what the instructions say, with all their fancy addressing modes, foo has to load two cache lines: one holding the shared_ptr and another holding the pointee data. If we instead passed bar* in a register, we'd need to grab only one cache line: the pointee's.

Sure. Maybe the caller already has a fully formed shared_ptr around somewhere but not in cache. Maybe foo often doesn't access the pointee. But how often does this situation arise?

[a_t48]

If you _maybe_ need to share ownership, the second is a little pessimistic - you always increase the ref count.

[Maxatar]

That is correct and I can see that being a justification for passing a const&, in fact the C++ Core Guidelines agree with you that such a scenario is the only acceptable reason for passing a shared_ptr by const&, although they encourage passing by value, or just passing a const T&.

[a_t48]

Obviously the correct way is to accept a templated type and use perfect forwarding. /s /s /s /s /s

[fooker]

Reference counting definitely slows down tight loops if you are not careful.

The way to avoid that in low latency code is to break the abstraction and operate with the raw pointer in the few areas where this could be a bottleneck.

It is usually not a bottleneck if your code is decently exploiting ipc, an extra addition or subtraction easily gets executed while some other operation is waiting a cycle for some cpu resource.

[shaklee3]

That's not true. It does slow things down because it has an atomic access. How slow depends on the platform.

unique_ptr does not slow things down.

[pjmlp]

C++ might have a bad reputation, but in many fields the only alternative, in terms of ecosystem, tooling and tribal knowledge is C.

Between those two, I rather pick the "Typescript for C" one.

[chipdart]

> I felt that C++ has a lot of bad publicity and if I want anyone to trust/try my code I would have to rebuild it in rust.

C++ gets bad publicity only from evangelists of the flavour of the month of self-described "successor of C++". They don't have a sales pitch beyond "C++ bad" and that's what they try to milk.

And yet the world runs on C++.

[rmlong]

std::shared_ptr definitely slows things down. It's non-intrusive therefore requires a memory indirection.