In my experience teaching undergrads they mostly get this stuff already. Their CompArch class has taught them the basics of branch prediction, cache coherence, and instruction caches; the trivial elements of performance.
I'm somewhat surprised the piece doesn't deal at all with a classic performance killer, false sharing, although it seems mostly concerned with single-threaded latency. The total lack of "free" optimization tricks like fat LTO, PGO, or even the standardized hinting attributes ([[likely]], [[unlikely]]) for optimizing icache layout was also surprising.
Neither this piece, nor my undergraduates, deal with the more nitty-gritty elements of performance. These mostly get into the usage specifics of particular IO APIs, synchronization primitives, IPC mechanisms, and some of the more esoteric compiler builtins.
Besides all that, what the nascent low-latency programmer almost always lacks, and the hardest thing to instill in them, is a certain paranoia. A genuine fear, hate, and anger, towards unnecessary allocations, copies, and other performance killers. A creeping feeling that causes them to compulsively run the benchmarks through callgrind looking for calls into the object cache that miss and go to an allocator in the middle of the hot loop.
I think a formative moment for me was when I was writing a low-latency server and I realized that constructing a vector I/O operation ended up being overall slower than just copying the small objects I was dealing with into a contiguous buffer and performing a single write. There's no such thing as a free copy, and that includes fat pointers.
> Fairly trivial base introduction to the subject.
Might be, but low-latency C++, in spite of being a field on its own, is a desert of information.
The best resources available at the moment on low-latency C++ are a hand full of lectures from C++ conferences which left much to be desired.
Putting aside the temptation to grandstand, this document is an outstanding contribution to the field and perhaps the first authoritative reference on the subject. Vague claims that you can piece together similar info from other courses does not count as a contribution, and helps no one.
> this document is an outstanding contribution to the field and perhaps the first authoritative reference on the subject.
I don't know how you arrive at this conclusion. The document really is an introduction to the same basic performance techniques that have been covered over and over. Loop unrolling, inlining, and the other techniques have appeared in countless textbooks and blog posts already.
I was disappointed to read the paper because they spent so much time covering really basic micro techniques but then didn't cover any of the more complicated issues mentioned in the parent comment.
I don't understand why you'd think this is an "outstanding contribution to the field" when it's basically a recap of simple techniques that have been covered countless times in textbooks and other works already. This paper may seem profound if someone has never, ever read anything about performance optimization before, but it's likely mundane to anyone who has worked on performance before or even wondered what inlining or -Funroll-loops does while reading some other code.
> A creeping feeling that causes them to compulsively run the benchmarks through callgrind
I'm happy I don't deal with such things these days, but I feel where the real paranoia always lies is the Heisenberg feeling of not even being able to even trust these things, the sneaky suspicion that the program is doing something different when I'm not measuring it.
On the software side I don't think HFT is as special a space as this paper makes it out to be.[1] Each year at cppcon there's another half-dozen talks going in depth on different elements of performance that cover more ground collectively than any single paper will.
Similarly, there's an immense amount of formal literature and textbooks out of the game development space that can be very useful to newcomers looking for structural approaches to high performance compute and IO loops. Games care a lot about local and network latency, the problem spaces aren't that far apart (and writing games is a very fun way to learn).
I don't have specific recommendations for holistic introductions to the field. I learn new techniques primarily through building things, watching conference talks, reading source code of other low latency projects, and discussion with coworkers.
[1]: HFT is quite special on the hardware side, which is discussed in the paper. The NICs, network stacks, and extensive integration of FPGAs do heavily differentiate the industry and I don't want to insinuate otherwise.
You will not find a lot of SystemVerilog programmers at a typical video game studio.
How I might approach it. Interested in feedback from people closer to the space.
First, split the load in to simple asset-specific data streams with a front-end FPGA for raw speed. Resist the temptation to actually execute here as the friction is too high for iteration, people, supply chain, etc. Input may be a FIX stream or similar, output is a series of asset-specific binary event streams along low-latency buses, split in to asset-specific segments of a scalable cluster of low-end MCUs. Second, get rid of the general purpose operating system assumption on your asset-specific MCU-based execution platform to enable faster turnaround using low-level code you can actually find people to write on hardware you can actually purchase. Third, profit? In such a setup you'd need to monitor the overall state with a general purpose OS based governor which could pause or change strategies by reprogramming the individual elements as required.
Just how low are the latencies involved? At a certain point you're better off paying to get the hardware closer to the core than bothering with engineering, right? I guess that heavily depends on the rules and available DCs / link infrastructure offered by the exchanges or pools in question. I would guess a number of profitable operations probably don't disclose which pools they connect to and make a business of front-running, regulations or terms of service be damned. In such cases, the relative network geographic latency between two points of execution is more powerful than the absolute latency to one.
The work I do is all in the single-to-low-double-digit microsecond range to give you an idea of timing constraints. I'm peripheral to HFT as a field though.
> First, split the load in to simple asset-specific data streams with a front-end FPGA for raw speed. Resist the temptation to actually execute here as the friction is too high for iteration, people, supply chain, etc.
This is largely incorrect, or more generously out-of-date, and it influences everything downstream of your explanation. Think of FPGAs as far more flexible GPUs and you're in the right arena. Input parsing and filtering are the obvious applications, but this is by no means the end state.
A wide variety of sanity checks and monitoring features are pushed to the FPGAs, fixed calculation tasks, and output generation. It is possible for the entire stack for some (or most, or all) transactions to be implemented at the FPGA layer. For such transactions the time magnitudes are mid-to-high triple digit nanoseconds. The stacks I've seen with my own two eyeballs talked to supervision algorithms over PCIe (which themselves must be not-slow, but not in the same realm as <10us work), but otherwise nothing crazy fancy. This is well covered in the older academic work on the subject [1], which is why I'm fairly certain its long out of date by now.
HRT has some public information on the pipeline they use for testing and verifying trading components implemented in HDL.[2] With the modern tooling, namely Verilator, development isn't significantly different than modern software development. If anything, SystemVerilog components are much easier to unit test than typical C++ code.
Beyond that it gets way too firm-specific to really comment on anything, and I'm certainly not the one to comment. There's maybe three dozen HFT firms in the entire United States? It's not a huge field with widely acknowledged industry norms.
> optimization tricks like fat LTO, PGO, or even the standardized hinting attributes ([[likely]], [[unlikely]]) for optimizing icache layout
If you do PGO, aren't hinting attributes counter-productive?
In fact, the common wisdom I mostly see compiler people express is that most of the time they're counter-productive even without PGO, and modern compilers trust their own analysis passes more than they trust these hints and will usually ignore them.
FWIW, the only times I've seen these hints in the wild were in places where the compiler could easily insert them, eg the null check after a malloc call.
I said "or even", if you're regularly using PGO they're irrelevant, but not everyone regularly uses PGO in a way that covers all their workloads.
The hinting attributes are exceptional for lone conditionals (not if/else trees) without obvious context to the compiler if it will frequently follow or skip the branch. Compilers are frequently conservative with such things and keep the code in the hot path.
The [[likely]] attribute then doesn't matter so much, but [[unlikely]] is absolutely respected and gets the code out of the hot path, especially with inlined into a large section. Godbolt is useful to verify this but obviously there's no substitute for benchmarking the performance impact.
> The output of this test is a test statistic (t-statistic) and an associated p-value. The t-statistic, also known as the score, is the result of the unit-root test on the residuals. A more negative t-statistic suggests that the residuals are more likely to be stationary. The p-value provides a measure of the probability that the null hypothesis of the test (no cointegration) is true. The results of your test yielded a p-value of approximately 0.0149 and a t-statistic of -3.7684.
I think they used an LLM to write this bit.
It's also a really weird example. They look at correlation of once-a-day close prices over five years, and then write code to calculate the spread with 65 microsecond latency. That doesn't actually make any sense as something to do. And you wouldn't be calculating statistics on the spread in your inner loop. And 65 microseconds is far too slow for an inner loop. I suppose the point is just to exercise some optimisation techniques - but this is a rather unrepresentative thing to optimise!
I've been looking to rebuild this in rust however. I reached the point where I implemented my own websocket protocol, authentication system, SSL etc. Then I realized that memory management and dependencies are a lot easier in rust. Especially for a one man software project.
It's not easy to get data structures like this right in C++. There are a couple of problems with your implementation of the queue.
Memory accesses can be reordered by both the compiler and the CPU, so you should use std::atomic for your producer and consumer positions to get the barriers described in the original LMAX Disruptor paper.
In the get method, you're returning a pointer to the element within the queue after bumping the consumer position (which frees the slot for the producer), so it can get overwritten while the user is accessing it.
And then your producer and consumer positions will most likely end up in the same cache line, leading to false sharing.
>> In the get method, you're returning a pointer to the element within the queue after bumping the consumer position (which frees the slot for the producer), so it can get overwritten while the user is accessing it. And then your producer and consumer positions will most likely end up in the same cache line, leading to false sharing.
I did not realize this. Thank you so much for pointing this out. I'm going to take a look.
>> use std::atomic for your producer
Yes, it is hard to get these data structures right. I used Martin Fowler's description of the LMAX algorithm which did not mention atomic. https://martinfowler.com/articles/lmax.html
I'll check out the paper.
T *item = &this->shared_mem_region
->entities[this->shared_mem_region->consumer_position];
this->shared_mem_region->consumer_position++;
this->shared_mem_region->consumer_position %= this->slots;
you can do this.
uint64_t mask = slot_count - 1; // all 1's in binary
item = &slots[ pos & mask ];
pos ++;
i.e. you can replace a division / modulo with a bitwise AND, saving a bit of computation. This requires that the size of the ringbuffer is a power of two.
What's more, you get to use sequence numbers over the full range of e.g. uint64_t. Wraparound is automatic. You can easily subtract two sequence numbers, this will work without a problem even accounting for wraparound. And you won't have to deal with stupid problems like having to leave one empty slot in the buffer because you would otherwise not be able to discern a full buffer from an empty one.
Naturally, you'll still want to be careful that the window of "live" sequence numbers never exceeds the size of your ringbuffer "window".
- 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.
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.
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.
The LMAX disruptor is a great data structure when you have threads bound to cores and most/all of them are uncontended. It has some terrible pathologies at the tails if you aren't using this pattern. Threads getting descheduled at bad times can really hurt.
SPSC ring buffers are going to be hard to beat for the system you are thinking of, and you can likely also implement work stealing using good old locks if you need it.
I think it made sense at the time. From what I understand, you can make Java run as fast as C++ if you're careful with it and use JIT. However, I have never tried such a thing and my source is hearsay from friends who have worked in financial institutions. Then you get added benefit of the Java ecosystem.
It took me about a year to build all of this stuff in C++. So I imagine since I've had to learn rust, it will probably take me the same amount of time if I can save time with dependencies.
The slide on measuring by having a fake server replaying order data, a second server calculating runtimes, the server under test, and a hardware switch to let you measure packet times is so delightfully hardcore.
I don't have any interest in working in finance, but it must be fun working on something so performance critical that buying a rack of hardware just for benchmarking is economically feasible.
I made a C++ logging library [1] that has many similarities to the LMAX disruptor. It appears to have found some use among the HFT community.
The original intent was to enable highly detailed logging without performance degradation for "post-mortem" debugging in production environments. I had coworkers who would refuse to include logging of certain important information for troubleshooting, because they were scared that it would impact performance. This put an end to that argument.
> The noted efficiency in compile-time dispatch is due to decisions about function calls being made during
the compilation phase. By bypassing the decision-making overhead present in runtime dispatch, programs
can execute more swiftly, thus boosting performance.
The other benefit with compile-time dispatch is that when the compiler can statically determine which function is being called, it may be able to inline the called function's code directly at the callsite. That eliminates all of the function call overhead and may also enable further optimizations (dead code elimination, constant propagation, etc.).
> That eliminates all of the function call overhead and may also enable further optimizations (dead code elimination, constant propagation, etc.).
AFAIK, the speedup is almost never function call overhead. As you mention at the tail end, it's all about the compiler optimizations being able to see past the dynamic branch. Good JITs support polymorphic inlining. My (somewhat dated) experience for C++ is that PGO is the solve for this, but it's not widely used. Instead people tend to avoid dynamic dispatch altogether in performance sensitive code.
I think the more general moral of the story is to avoid all kinds of unnecessary dynamic branching in hot sections of code in any language unless you have strong/confidence your compiler/JIT is seeing through it.
The real performance depends on the runtime behavior of the machine as well as compiler optimizations. I thought this talk was very interesting on this subject.
Is there any good reason for high-frequency trading to exist? People often complain about bitcoin wasting energy, but oddly this gets a free pass despite this being a definite net negative to society as far as I can tell.
Bid/ask spreads are far narrower than they were previously. If you look at the profits of the HFT industry as a whole they aren't that large (low billions) and their dollar volume is in the trillions. Hard to argue that the industry is wildly prosocial but making spreads narrower does mean less money goes to middlemen.
> but making spreads narrower does mean less money goes to middlemen.
On individual trades. I would think you'd have to also argue that their high overall trading volume is somehow also a benefit to the broader market or at the very least that it does not outcompete the benefits of narrowing.
I would argue that HFT is a rather small space, albeit pretty concentrated. It's several orders of magnitude smaller in terms of energy wasting than Bitcoin.
The only positive from HFT is liquidity and tighter spreads, but it also depends what people put into HFT definition. For example, Robinhood and free trading, probably wouldn't exist without it.
They are taking a part of the cake that previously went to brokers and banks. HFT is not in a business of screwing 'the little guy'.
From my perspective there is little to none negative to the society. If somebody is investing long term in the stock market, he couldn't care less about HFT.
Buying and quickly selling a stock requires a buyer and a seller, so it doesn't seem like an intrinsic property that real liquidity would be increased. Moreover, fast automated trading seems likely to increase volatility.
I tend to agree that for long term investment it probably doesn't make a huge difference except for possible cumulative effects of decreased liquidity, increased volatility, flash crashes, etc. Also possibly a loss of small investor confidence since the game seems even more rigged in a way that they cannot compete with.
Warren Buffett proposed that the stock market should be open less frequently, like once a quarter or similar. This would encourage long-term investing rather than reacting to speculation.
Regardless, there are no natural events that necessitate high-frequency trading. The underlying value of things rarely changes very quickly, and if it does it's not volatile, rather it's a firm transiton.
> Warren Buffett proposed that the stock market should be open less frequently
This will result in another market where deals will be made and then finalized on that 'official' when it opens. It's like with employee stock. You can sell it before you can...
AIUI the point of HFT isn't to trade frequently, but rather to change offer/bid prices in the smallest possible steps (small along both axes) while waiting for someone to accept the offer/bid.
Why shouldn’t people be allowed to do both? I don’t see much of an advantage to making the markets less agile.
It would be nice to be able to buy and sell stocks more than once a quarter, especially given plenty of events that do affect the perceived value of a company happen more frequently than that
The value of a stock usually doesn't change every millisecond. There doesn't seem to be a huge public need to pick winners and losers based on ping time.
Seems like a testable hypothesis. Choose the four times a year that the stock is at a fair price, and buy when it goes below it and sell when it goes above it?
Non-bitcoin transactions are just a couple of entries in various databases. Mining bitcoin is intense number crunching.
HFT makes the financial markets a tiny bit more accurate by resolving inconsistencies (for example three pairs of currencies can get out of whack with one another) and obvious mispricings (for various definitions of "obvious")
That's a nice fairy tale that they probably tell their kids when asked, but what the profitable firms are doing at the cutting edge is inducing responses in the other guys' robots, in a phase that the antagonist controls, then trading against what they know is about to happen. It is literally market manipulation. A way to kill off this entire field of endeavor is to charge a tax on cancelled orders.
How far have you tried to tell, and do you buy/sell stocks?
There's someone on the other side of your trade when you want to trade something. You're more likely than not choosing to interact with an HFT player at your price. If you're getting a better price, that's money that you get to keep.
*I'm going to disagree on "free pass" also. HFT is pretty often criticized here.
- Increased liquidity. Ensures there's actually something to be traded available globally, and swiftly moves it to places where it's lacking.
- Tighter spreads, the difference between you buying and then selling again is lower. Which often is good for the "actual users" of the market.
- Global prices / less geographical differences in prices. Generally you can trust you get the right price no matter what venue you trade at, as any arbitrage opportunity has likely already been executed on.
> Tighter spreads, the difference between you buying and then selling again is lower. Which often is good for the "actual users" of the market.
I just wanted to highlight this one in particular - the spread is tighter because HFTs eat the spread and reduce the error that market players can benefit from. The spread is disappearing because of rent-seeking from the HFTs.
Yes: it put a much larger, more expensive, and less efficient part of wall street out of business. Before it was done with computers, it was done with lots of people doing the same job. What was that job? Well, if you want to go to a market and sell something, you generally would like for there to be someone there who's buying it. But it's not always the case that there's a buyer there right at the time who actually wants the item for their own use. The inverse is also true for a prospective buyer. Enter middle-men or market makers who just hang around the marketplace, learning roughly how much people will buy or sell a given good for, and buy it for slightly less than they can sell it later for. This is actually generally useful if you just want to buy or sell something.
Now, does this need to get towards milli-seconds or nano-seconds? No, this is just the equivalent of many of these middle-men racing to give you an offer. But it's (part of) how they compete with each other, and as they do so they squeeze the margins of the industry as a whole: In fact the profits of HFT firms have decreased as a percentage of the overall market and in absolute terms after the initial peak as they displaced the day traders doing the same thing.
> it's not always the case that there's a buyer there right at the time
This hits the nail on the head. For a trade to happen, counterparties need to meet in price and in time. A market place is useless if there is nobody around to buy or sell at the same time you do.
The core service market makers provide is not liquidity. It's immediacy: they offer (put up) liquidity in order to capture trades, but the value proposition for other traders - and the exchanges! - is that there is someone to take the other side of a trade when a non-MM entity wants to buy or sell instruments.
It took me a long time to understand what the difference is. And in order to make sure that there is sufficient liquidity in place, exchanges set up both contractual requirements and incentive structures for their market makers.
Central counterparty concept implemented by most exchanges is a valid service, as otherwise counterparty risk management would be a nightmare - an example of a useful “middleman”.
Under the carpet deals would make it less transparent. It would be hard to detect that top manager sold all his stock to competitors and now is making decisions in their favor.
You mean like settling trades every 0.25 seconds or something like that? Wouldn't there be a queue of trades piling up every 0.25 seconds, incentivizing maximum speed anyway?
If you don't like it, try to convince regulators that it shouldn't be legal and provide a framework for criminalizing/fining it without unintended consequences, and then find a way to pay regulators more in bribes than the HFT shops do, even though their pockets are deeper than yours, and then things may change.
If that sounds impossible, that's another answer to your question
> a definite net negative to society as far as I can tell.
What did you examine to reach that conclusion? If high-frequency trading were positive for society, what would you expect to be different?
The reason for high-frequency trading to exist is that the sub-penny rule makes it illegal to compete on price so you have to compete on speed instead. Abolishing the sub-penny rule would mean high-frequency trading profits got competed-away to nothing, although frankly they're already pretty close. The whole industry is basically an irrelevant piece of plumbing anyway.
> A net negative to society, but a positive for the wealthiest.
No.
When your passive index fund manager rebalances every month because “NVDA is now overweighted in VTI, QQQ” the manager does not care about the bid/ask spread.
When VTI is $1.6 trillion, even a $0.01 difference in price translates to a loss $60 million for the passive 401k, IRA, investors.
HFT reduces the bid/ask spread, and “gives this $60 million back” to the passive investors for every $0.01 price difference, every month. Note that VTI mid price at time of writing is $272.49.
Readit News
In my experience teaching undergrads they mostly get this stuff already. Their CompArch class has taught them the basics of branch prediction, cache coherence, and instruction caches; the trivial elements of performance.
I'm somewhat surprised the piece doesn't deal at all with a classic performance killer, false sharing, although it seems mostly concerned with single-threaded latency. The total lack of "free" optimization tricks like fat LTO, PGO, or even the standardized hinting attributes ([[likely]], [[unlikely]]) for optimizing icache layout was also surprising.
Neither this piece, nor my undergraduates, deal with the more nitty-gritty elements of performance. These mostly get into the usage specifics of particular IO APIs, synchronization primitives, IPC mechanisms, and some of the more esoteric compiler builtins.
Besides all that, what the nascent low-latency programmer almost always lacks, and the hardest thing to instill in them, is a certain paranoia. A genuine fear, hate, and anger, towards unnecessary allocations, copies, and other performance killers. A creeping feeling that causes them to compulsively run the benchmarks through callgrind looking for calls into the object cache that miss and go to an allocator in the middle of the hot loop.
I think a formative moment for me was when I was writing a low-latency server and I realized that constructing a vector I/O operation ended up being overall slower than just copying the small objects I was dealing with into a contiguous buffer and performing a single write. There's no such thing as a free copy, and that includes fat pointers.
Might be, but low-latency C++, in spite of being a field on its own, is a desert of information.
The best resources available at the moment on low-latency C++ are a hand full of lectures from C++ conferences which left much to be desired.
Putting aside the temptation to grandstand, this document is an outstanding contribution to the field and perhaps the first authoritative reference on the subject. Vague claims that you can piece together similar info from other courses does not count as a contribution, and helps no one.
I don't know how you arrive at this conclusion. The document really is an introduction to the same basic performance techniques that have been covered over and over. Loop unrolling, inlining, and the other techniques have appeared in countless textbooks and blog posts already.
I was disappointed to read the paper because they spent so much time covering really basic micro techniques but then didn't cover any of the more complicated issues mentioned in the parent comment.
I don't understand why you'd think this is an "outstanding contribution to the field" when it's basically a recap of simple techniques that have been covered countless times in textbooks and other works already. This paper may seem profound if someone has never, ever read anything about performance optimization before, but it's likely mundane to anyone who has worked on performance before or even wondered what inlining or -Funroll-loops does while reading some other code.
I'm happy I don't deal with such things these days, but I feel where the real paranoia always lies is the Heisenberg feeling of not even being able to even trust these things, the sneaky suspicion that the program is doing something different when I'm not measuring it.
Similarly, there's an immense amount of formal literature and textbooks out of the game development space that can be very useful to newcomers looking for structural approaches to high performance compute and IO loops. Games care a lot about local and network latency, the problem spaces aren't that far apart (and writing games is a very fun way to learn).
I don't have specific recommendations for holistic introductions to the field. I learn new techniques primarily through building things, watching conference talks, reading source code of other low latency projects, and discussion with coworkers.
[1]: HFT is quite special on the hardware side, which is discussed in the paper. The NICs, network stacks, and extensive integration of FPGAs do heavily differentiate the industry and I don't want to insinuate otherwise.
You will not find a lot of SystemVerilog programmers at a typical video game studio.
The author has a strong game development (engine and tooling) background and I have found it incredibly useful.
It also satisfies the requirement for "A genuine fear, hate, and anger, towards unnecessary allocations, copies, and other performance killers."
First, split the load in to simple asset-specific data streams with a front-end FPGA for raw speed. Resist the temptation to actually execute here as the friction is too high for iteration, people, supply chain, etc. Input may be a FIX stream or similar, output is a series of asset-specific binary event streams along low-latency buses, split in to asset-specific segments of a scalable cluster of low-end MCUs. Second, get rid of the general purpose operating system assumption on your asset-specific MCU-based execution platform to enable faster turnaround using low-level code you can actually find people to write on hardware you can actually purchase. Third, profit? In such a setup you'd need to monitor the overall state with a general purpose OS based governor which could pause or change strategies by reprogramming the individual elements as required.
Just how low are the latencies involved? At a certain point you're better off paying to get the hardware closer to the core than bothering with engineering, right? I guess that heavily depends on the rules and available DCs / link infrastructure offered by the exchanges or pools in question. I would guess a number of profitable operations probably don't disclose which pools they connect to and make a business of front-running, regulations or terms of service be damned. In such cases, the relative network geographic latency between two points of execution is more powerful than the absolute latency to one.
> First, split the load in to simple asset-specific data streams with a front-end FPGA for raw speed. Resist the temptation to actually execute here as the friction is too high for iteration, people, supply chain, etc.
This is largely incorrect, or more generously out-of-date, and it influences everything downstream of your explanation. Think of FPGAs as far more flexible GPUs and you're in the right arena. Input parsing and filtering are the obvious applications, but this is by no means the end state.
A wide variety of sanity checks and monitoring features are pushed to the FPGAs, fixed calculation tasks, and output generation. It is possible for the entire stack for some (or most, or all) transactions to be implemented at the FPGA layer. For such transactions the time magnitudes are mid-to-high triple digit nanoseconds. The stacks I've seen with my own two eyeballs talked to supervision algorithms over PCIe (which themselves must be not-slow, but not in the same realm as <10us work), but otherwise nothing crazy fancy. This is well covered in the older academic work on the subject [1], which is why I'm fairly certain its long out of date by now.
HRT has some public information on the pipeline they use for testing and verifying trading components implemented in HDL.[2] With the modern tooling, namely Verilator, development isn't significantly different than modern software development. If anything, SystemVerilog components are much easier to unit test than typical C++ code.
Beyond that it gets way too firm-specific to really comment on anything, and I'm certainly not the one to comment. There's maybe three dozen HFT firms in the entire United States? It's not a huge field with widely acknowledged industry norms.
[1]: https://ieeexplore.ieee.org/document/6299067
[2]: https://www.hudsonrivertrading.com/hrtbeat/verify-custom-har...
If you do PGO, aren't hinting attributes counter-productive?
In fact, the common wisdom I mostly see compiler people express is that most of the time they're counter-productive even without PGO, and modern compilers trust their own analysis passes more than they trust these hints and will usually ignore them.
FWIW, the only times I've seen these hints in the wild were in places where the compiler could easily insert them, eg the null check after a malloc call.
The hinting attributes are exceptional for lone conditionals (not if/else trees) without obvious context to the compiler if it will frequently follow or skip the branch. Compilers are frequently conservative with such things and keep the code in the hot path.
The [[likely]] attribute then doesn't matter so much, but [[unlikely]] is absolutely respected and gets the code out of the hot path, especially with inlined into a large section. Godbolt is useful to verify this but obviously there's no substitute for benchmarking the performance impact.
Please elaborate on those other performance killers.
> The output of this test is a test statistic (t-statistic) and an associated p-value. The t-statistic, also known as the score, is the result of the unit-root test on the residuals. A more negative t-statistic suggests that the residuals are more likely to be stationary. The p-value provides a measure of the probability that the null hypothesis of the test (no cointegration) is true. The results of your test yielded a p-value of approximately 0.0149 and a t-statistic of -3.7684.
I think they used an LLM to write this bit.
It's also a really weird example. They look at correlation of once-a-day close prices over five years, and then write code to calculate the spread with 65 microsecond latency. That doesn't actually make any sense as something to do. And you wouldn't be calculating statistics on the spread in your inner loop. And 65 microseconds is far too slow for an inner loop. I suppose the point is just to exercise some optimisation techniques - but this is a rather unrepresentative thing to optimise!
And a basic implementation of the LMAX disruptor as a couple C++ files https://github.com/sneilan/lmax-disruptor-tutorial
I've been looking to rebuild this in rust however. I reached the point where I implemented my own websocket protocol, authentication system, SSL etc. Then I realized that memory management and dependencies are a lot easier in rust. Especially for a one man software project.
I did not realize this. Thank you so much for pointing this out. I'm going to take a look.
>> use std::atomic for your producer
Yes, it is hard to get these data structures right. I used Martin Fowler's description of the LMAX algorithm which did not mention atomic. https://martinfowler.com/articles/lmax.html I'll check out the paper.
What's more, you get to use sequence numbers over the full range of e.g. uint64_t. Wraparound is automatic. You can easily subtract two sequence numbers, this will work without a problem even accounting for wraparound. And you won't have to deal with stupid problems like having to leave one empty slot in the buffer because you would otherwise not be able to discern a full buffer from an empty one.
Naturally, you'll still want to be careful that the window of "live" sequence numbers never exceeds the size of your ringbuffer "window".
- 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.
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.
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.
SPSC ring buffers are going to be hard to beat for the system you are thinking of, and you can likely also implement work stealing using good old locks if you need it.
https://martinfowler.com/articles/lmax.html
The slide on measuring by having a fake server replaying order data, a second server calculating runtimes, the server under test, and a hardware switch to let you measure packet times is so delightfully hardcore.
I don't have any interest in working in finance, but it must be fun working on something so performance critical that buying a rack of hardware just for benchmarking is economically feasible.
But of course you don't have to buy a rack of servers for testing, you can rent it. Servers are a quickly depreciating asset, why invest in them?
The original intent was to enable highly detailed logging without performance degradation for "post-mortem" debugging in production environments. I had coworkers who would refuse to include logging of certain important information for troubleshooting, because they were scared that it would impact performance. This put an end to that argument.
[1] https://github.com/mattiasflodin/reckless
The other benefit with compile-time dispatch is that when the compiler can statically determine which function is being called, it may be able to inline the called function's code directly at the callsite. That eliminates all of the function call overhead and may also enable further optimizations (dead code elimination, constant propagation, etc.).
AFAIK, the speedup is almost never function call overhead. As you mention at the tail end, it's all about the compiler optimizations being able to see past the dynamic branch. Good JITs support polymorphic inlining. My (somewhat dated) experience for C++ is that PGO is the solve for this, but it's not widely used. Instead people tend to avoid dynamic dispatch altogether in performance sensitive code.
I think the more general moral of the story is to avoid all kinds of unnecessary dynamic branching in hot sections of code in any language unless you have strong/confidence your compiler/JIT is seeing through it.
https://youtu.be/i5MAXAxp_Tw
Though my impression is that compilers tend to be fairly conservative about inlining so that don't risk the inlining being a pessimization.
On individual trades. I would think you'd have to also argue that their high overall trading volume is somehow also a benefit to the broader market or at the very least that it does not outcompete the benefits of narrowing.
I would argue that HFT is a rather small space, albeit pretty concentrated. It's several orders of magnitude smaller in terms of energy wasting than Bitcoin.
The only positive from HFT is liquidity and tighter spreads, but it also depends what people put into HFT definition. For example, Robinhood and free trading, probably wouldn't exist without it.
They are taking a part of the cake that previously went to brokers and banks. HFT is not in a business of screwing 'the little guy'.
From my perspective there is little to none negative to the society. If somebody is investing long term in the stock market, he couldn't care less about HFT.
I tend to agree that for long term investment it probably doesn't make a huge difference except for possible cumulative effects of decreased liquidity, increased volatility, flash crashes, etc. Also possibly a loss of small investor confidence since the game seems even more rigged in a way that they cannot compete with.
Regardless, there are no natural events that necessitate high-frequency trading. The underlying value of things rarely changes very quickly, and if it does it's not volatile, rather it's a firm transiton.
This will result in another market where deals will be made and then finalized on that 'official' when it opens. It's like with employee stock. You can sell it before you can...
It would be nice to be able to buy and sell stocks more than once a quarter, especially given plenty of events that do affect the perceived value of a company happen more frequently than that
HFT makes the financial markets a tiny bit more accurate by resolving inconsistencies (for example three pairs of currencies can get out of whack with one another) and obvious mispricings (for various definitions of "obvious")
There's someone on the other side of your trade when you want to trade something. You're more likely than not choosing to interact with an HFT player at your price. If you're getting a better price, that's money that you get to keep.
*I'm going to disagree on "free pass" also. HFT is pretty often criticized here.
- Increased liquidity. Ensures there's actually something to be traded available globally, and swiftly moves it to places where it's lacking.
- Tighter spreads, the difference between you buying and then selling again is lower. Which often is good for the "actual users" of the market.
- Global prices / less geographical differences in prices. Generally you can trust you get the right price no matter what venue you trade at, as any arbitrage opportunity has likely already been executed on.
- etc..
I just wanted to highlight this one in particular - the spread is tighter because HFTs eat the spread and reduce the error that market players can benefit from. The spread is disappearing because of rent-seeking from the HFTs.
Now, does this need to get towards milli-seconds or nano-seconds? No, this is just the equivalent of many of these middle-men racing to give you an offer. But it's (part of) how they compete with each other, and as they do so they squeeze the margins of the industry as a whole: In fact the profits of HFT firms have decreased as a percentage of the overall market and in absolute terms after the initial peak as they displaced the day traders doing the same thing.
This hits the nail on the head. For a trade to happen, counterparties need to meet in price and in time. A market place is useless if there is nobody around to buy or sell at the same time you do.
The core service market makers provide is not liquidity. It's immediacy: they offer (put up) liquidity in order to capture trades, but the value proposition for other traders - and the exchanges! - is that there is someone to take the other side of a trade when a non-MM entity wants to buy or sell instruments.
It took me a long time to understand what the difference is. And in order to make sure that there is sufficient liquidity in place, exchanges set up both contractual requirements and incentive structures for their market makers.
Deleted Comment
Because it's legal and profitable.
If you don't like it, try to convince regulators that it shouldn't be legal and provide a framework for criminalizing/fining it without unintended consequences, and then find a way to pay regulators more in bribes than the HFT shops do, even though their pockets are deeper than yours, and then things may change.
If that sounds impossible, that's another answer to your question
What did you examine to reach that conclusion? If high-frequency trading were positive for society, what would you expect to be different?
The reason for high-frequency trading to exist is that the sub-penny rule makes it illegal to compete on price so you have to compete on speed instead. Abolishing the sub-penny rule would mean high-frequency trading profits got competed-away to nothing, although frankly they're already pretty close. The whole industry is basically an irrelevant piece of plumbing anyway.
No.
When your passive index fund manager rebalances every month because “NVDA is now overweighted in VTI, QQQ” the manager does not care about the bid/ask spread.
When VTI is $1.6 trillion, even a $0.01 difference in price translates to a loss $60 million for the passive 401k, IRA, investors.
HFT reduces the bid/ask spread, and “gives this $60 million back” to the passive investors for every $0.01 price difference, every month. Note that VTI mid price at time of writing is $272.49.
https://github.com/CppCon/CppCon2017/tree/master/Presentatio...
and up