This seems old (2013ish?). There're newer "key/value on FPGA" papers that're more modern.
If memcached papers have one thing in common, it's an uncanny ability to get the comparison software to run as slowly as possible. 100k ops/sec/core is what you get when using a single client connection with blocking I/O. Using more clients (as per a normal prod setup) or pipelining queries is more like 1m ops+/core, with writes scaling worse than reads. In production it's easy to get some level of pipelining (multigets, clustered keys, etc), since you're rarely just fetching a key and then blocking.
A much better FPGA paper would show scales of at what level the syscall overhead becomes most of the CPU usage, as well as any measured latency improvements. I think some of the other papers address latency at least.
In reality it hardly matters. If you're hitting memcached hard enough with tiny values for it to matter, ensuring keys are clustered and pipelined is a lot less maintenance overhead than deploying FPGA's.
Can confirm this eerie ability of FPGA and h/w folks from another domain (regular expressions - I'm the designer of Hyperscan, a s/w solution).
Every corner case that could be found in s/w was always the topic of a excited benchmark. Also, the old trick of 'hey, let's drop all the matches on the floor in our h/w or FPGA, while getting a huge number of matches in s/w and making the s/w guys look ridiculous'.
Every time I read a paper touting a great new speedup on FPGA (over some crap s/w implementation) I'm reminded of that old joke about the Texan visiting Israel and telling the owner of some small farm that "he can get on a tractor and ride for days without getting the the boundary of his property." The Israeli nods sympathetically and says "Yes, I too used to have a tractor like that".
Especially when we have Xeon -D that goes up to 16 Core.
AMD EPYC that gives more Core per dollar. And more IPC soon in Zen 2. 7nm and 10nm from AMD and Intel next year on Server. Not to mention it now Support ARMv8
So Excuse my ignorance, apart from AWS or Azure Scales, what would anyone uses Memcached on FPGA?
* I wouldn't mind if the system was simple plug and play and has all the benefits, cost saving without the headache. But very rarely are any technology deployment without any headache or hassle free.
Our startup is working on accelerators using FPGAs on AWS including memcached.
Using a single AWS F1 (FPGA) instance, our Memcached accelerator achieves over 11 million ops/sec at less than 300 microsecond latency. Compared to ElastiCache, the AWS-managed CPU Memcached server, our Memcached accelerator offers 9X better throughput, 9X lower latency, and 10X better throughput/$.
I am the co-founder of Plunify. We have an ML software solution (InTime) that optimizes FPGA design by tuning the parameters of the compilation, i.e. synthesis, P&R. I don't know much about memcached accelerators but if it is a performance driven application, like HFT, I believe we can make it go even faster, i.e. increase the FMax. I read from your website that you are using Intel PSG devices. We often see designers using seeds to close timing or optimize the timing, but that is leaving performance on the table. For more details: https://support.plunify.com/en/2018/04/17/compare-timing-per...
Happy to talk more at kirvy@plunify.com if you are interested. Congrats on getting a seed round from Intel Capital!
Interesting, what synthesis settings have you found have the most impact? I have also seen FPGA designers trying different seeds when closing timing.
In this case, AWS provides an FPGA shell for external interfaces that has a maximum clock frequency of 250MHz. We have been able to meet this timing constraint without many issues. But we will keep you in mind for the Intel FPGA boards we are working with now.
what synthesis settings have you found have the most impact?
The answer is - it depends. Unfortunately, we have not found a "golden" combination of settings yet. If you have a highly congested design, synthesis does help a lot, but not all cases. There are correlations between the settings, so if A is good, B is good, A+B could be bad. Seeds belong to a category of techniques that we classify as random. For example, for Vivado, although Xilinx removed the seeds feature, we created a technique to trigger randomness in the placement using a property of Vivado.
What we do is not new in the sense that settings exploration has always been around. But with cloud compute resources and ML approaches, it really enables timing closure and optimization methods in a cheaper and more disciplined fashion.
We are also very interested in users of OpenCL/HLS/C. The translated RTL is often not as optimized/readable as what an RTL designer will do if he/she does it directly in RTL. Our tool (InTime) can be a good boost to the performance of such RTL.
How is elasticache so slow? what instances does it run on?
edit: r4.4xlarge as per the link. 16vcpu? You should be able to beat on latency but beating on throughput means elasticache is misconfigured, likely. Or you're putting on way too much set traffic (think I saw you set the bench to 1:1 ratio of gets to sets?)
I wouldn't characterize Elasticache as running slow, a single instance in this case is handling 1.3M request/sec. But we can be 9X faster by batching multiple requests per packet and then offloading the TCP network stack and memcached operations to the FPGA. The FPGA allows us to handle the requests at network line-rate, even with small 100-byte requests. On Elasticache, past a certain point these small requests start to overload the CPU.
The interesting part is the FPGA could still do much more computation (for example, compression or encryption) while maintaining the same throughput due to hardware pipelining. We described this concept further in the blog post I linked to.
I characterize it as slow because I know it can saturate the packet rate AWS gives it with software memcached. If the packet rate were much higher then you might win out.
The only reason why you can claim 9x latency is because you've saturated the worker threads. You should still win on latency even if it were properly bottlenecking on the network, but 9x throughput and 9x latency is completely false as a capacity limit in this test.
The other issue is 100 bytes isn't typical. It's common but almost every user has a varied workload. Deploying FPGA's for the larger cache values ends up being a waste. I designed a new storage system based off of offloading larger cold keys to flash, even.
Do you have a source showing elasticache running faster than this? For example, Redis labs was only able to achieve 10M req/sec by using 6 m4.16xlarge instances which are double the price of the CPU instance we used:
https://dzone.com/articles/10m-opssec-1msec-latency-with-onl...
100-500 byte values are the majority of requests at companies like Facebook and Lyft for their key value clusters. For large value sizes the network interface becomes the bottleneck so FPGAs won’t be able to help.
I wrote the current version of OSS memcached. I don't know how elasticache is configured, but as I said memcached itself can definitely saturate the network from that instance. Either the version they run is too old or it's misconfigured. If I were to compare a "custom FPGA caching service" vs something memcached like, I would take the same 4xlarge instance and just run memcached on it.
On a large enough machine I've gotten it up past 55 million read ops/sec. It's quite good at read throughput.
I'm also familiar with the cache clusters at major companies.
Although there are Ethernet transceivers on the AWS FPGAs (these are Xilinx UltraScale+ VU9P FPGAs) they are unused and not connected directly to the AWS network. Instead the FPGAs are connected over PCIe to a host server, which has a standard NIC. This required us to use DPDK to bypass the kernel and pass raw network traffic directly to/from the FPGA over PCIe.
We have been told by cloud providers that the FPGA cannot be directly connected due to network security concerns. Since there is no easy way to control how the arbitrary hardware programmed by users on the FPGA will interact with their network. Microsoft has been using FPGAs directly connected to their network (called a bump-in-the-wire architecture) for the FPGAs used in their datacenter (see Project Catapult for details). But these FPGAs are not programmable by Azure users yet.
I wonder if you could bound the problem with formal verification so that a custom design could only expose a certain activity profile, or an embedded hardware kill switch that would eliminate access after some detectable malicious interaction (at the hardware level).
Would that even be possible, or does it boil down to the halting problem (how can you know a sequence of transactions is malicious)?
I'm completely naive, but even so I'm tentatively confident saying yes, it's a halting problem thing.
My understanding of formal verification is a) that it doesn't eliminate bugs, it just reduces their probability by a percentage, and b) that it's academically very attractive (which is why HN is constantly hearing about it) but practically very expensive in terms of investment/return, in the sense that formal verification requires formal modeling and so forth in order to exist in the first place, and that requires actual design lockdown, and suddenly everything got Complicated™ :)
You could possibly get away with an FPGA-targeting compiler (low-relevance related conversation: https://news.ycombinator.com/item?id=17151024) set up in a stripped-down form, but you'd need to lock it down a bit tighter than https://www.corsix.org/content/malicious-luajit-bytecode (a bit old, but illustrative of a point), and then the question would be whether that would negate the performance benefit of the FPGA in the first place.
About to congratulate you that you are about to make competition to Alibaba.com.
On our side, maxing the IOPS was the easy part. The hard one was to marry the protocol with converged/deterministic Ethernet with RDMA. We were split in between "one request, one frame/packet burst" vs "all requests are somehow smartly aligned with frames by stateful logic." The first one was surprisingly susceptible to performance artifacts due to varying round trip latency of few microseconds, thus it was possible to get packets in transit being dropped due to receiving NIC (a top tier hardware) being momentarily overloaded.
You have an advantage of being DC provider independent, and can jump the AWS ship whenever you want. Alibaba's solution will be tied to its infrastructure with its very expensive RDMA capable network.
Sounds interesting, are there any papers or public details that you could point me to with more technical information about this Alibaba RDMA-based memcached project? Alibaba also has FPGA instances available and we have been investigating their cloud offering.
Invite only till at least autumn, and they have no English speaking support (it is for their Chinese only DCs.)
Google Alibaba China F1 or F2.
As for the offering, the idea is that end the clients will only have to deal with SDK and libs on instances, not raw RDMA or anything related to internal infrastructure. This is how much I am allowed to say besides the fact of its existence.
From their current experience, and that of other hosting providers, not many people who go with F1, F2, or other FPGA instance actually do reap any benefit, and some drop mid-way. That's why they want to get more people using them pass the "toying with it stage." The "Herokuification" (god, Heroku sounds beyond hilarious in Russian) is there to let people use the common APIs while getting benefit from FPGA performance, without dealing with things outside of average webdev area of expertise.
If memcached papers have one thing in common, it's an uncanny ability to get the comparison software to run as slowly as possible. 100k ops/sec/core is what you get when using a single client connection with blocking I/O. Using more clients (as per a normal prod setup) or pipelining queries is more like 1m ops+/core, with writes scaling worse than reads. In production it's easy to get some level of pipelining (multigets, clustered keys, etc), since you're rarely just fetching a key and then blocking.
A much better FPGA paper would show scales of at what level the syscall overhead becomes most of the CPU usage, as well as any measured latency improvements. I think some of the other papers address latency at least.
In reality it hardly matters. If you're hitting memcached hard enough with tiny values for it to matter, ensuring keys are clustered and pipelined is a lot less maintenance overhead than deploying FPGA's.