[dormando]

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.

[andrewcanis]

Our assumption was that elasticache would be highly optimized by Amazon. Remember that these are virtual machines which means limitations such as packet per second throttling. What specific configuration options do you think are missing?

In the latest version of memcached have you added support for batching/pipelining multiple requests per packet? Because this was crucial for achieving high requests/sec in this example.

Were the 55M requests/sec coming from another machine? Even with small 100B values you would need a minimum of a 44 Gbps network link. How many cores were required? In our benchmark we wanted a fair comparison between instances of similar price and RAM size.

[dormando]

They stopped updating it a few years ago; it's probably also not as well tuned as you think. I'd need to see the output of "stats settings" from a running instance to know for sure. I also have no idea if it's a hacked fork or not.

Odds are pretty good it's left at the default of 4 worker threads... so on a 16 vcpu instance that's not going to reach great heights. Since it's a 1.4.x version (years old), it's missing some newer features that both help in average latency and memory efficiency. Or rather, a lot of them are there but disabled by default.

Memcached has allowed pipelining since it was created. For the ASCII protocol, packing multiple responses into single packets is done via a straight multiget. You can send multiple requests in a single packet for any protocol and any command.

My stress utility (https://github.com/memcached/mc-crusher) has options for pipelining requests, and using multigets ascii packed get responses. I test to the limit of lock scaling for each individual subsystem.

The 55M test required running mc-crusher via localhost, there's no network that can go that fast. My point is you're limited by the network throughput, not the CPU. In that particular 55M test, all cores were used, but ~7-8 of them were used by mc-crusher... so the real limit for the machine is even higher. It did have a lot of cores. 48ish?

You can still do apples/apples with instance sizes... but given everything I know about this thing, unless those cores are extremely slow, hitting 11m ops/sec shouldn't be an issue. Or at least, with minimal fiddling it should hit 6-8m, which doesn't give you a crazy 9x figure.

You do need to stop doing 1:1 get/set ratio though. Sets don't scale very well since I've generally never had complaints about the speed. I'd say a highly conservative test would be 5:1 get/set. Production workloads are typically even higher than that. (that said I do intend to speed them up more, it's just lowish priority.. the LRU locks are highly granular, so spreading sets across different slab classes can help mutation perf a lot).

[andrewcanis]

I'm still seeing similar results (~1M req/sec) after compiling your latest version of memcached from github and running with 16 worker threads. I just spun up two r4.4xlarge instances (one for client and one for the memcached server). I'm using memtier_benchmark with pipelining of 16 requests, 100B values, 10:1 get/set ratio. I compiled mc-crusher but you'll have to let me know the command to run because the readme wasn't clear.

One main constraint here is that we are using AWS virtual machine instances on the cloud. My guess is your previous experience is with physical servers. The FPGA performance is also significantly better when you can use the physical board with a direct ethernet connection, pipelining isn't required in this case the FPGA can handle minimum sized ethernet packets at line rate.

Another question, in your experience is compression/encryption used much with memcached? Because this is another area where the FPGA can compute much faster.

[dormando]

mis-threaded my response below (didn't have a reply button?), so see that too.

Just signed up for a personal AWS account and manually started an r4.4xlarge for target and c5.4xlarge for source (same CPU's and networking capability?, but it wasn't allowing me to just start two r4.4xlarge...).

got it up to 15M hits/sec for pure mget test.

results: https://gist.github.com/dormando/910134e85279710b970bd2c8af8...

[andrewcanis]

Thanks for the details on how to use your benchmark script and for taking the time to investigate this. I hadn’t heard of your benchmark before and mc-crusher seems to work a bit differently than memtier_benchmark.

First a few significant differences:

1) Your value size is 10B which completely changes the results. Let’s keep the value size at 100B, which is more realistic.

2) The ratio of gets to sets significantly affects the requests per sec. We were assuming 1:1 ratio when we did our measurements. Increasing the percentage of gets really speeds up req/sec. We didn’t observe this effect on elasticache. Is this a recent improvement in the github version of memcached?

3) Your benchmark is using multiple keys in the same get command. What memtier does is pipeline multiple get commands each with one key. This seems more realistic.

4) We pipelined 16 get commands per packet while your configuration had 50 keys per get command.

I was able to reproduce the same setup as we had with ~1.2M req/sec with your mc-crusher benchmark using the following config. This has 1:1 get to set ratio with pipeline 16 and value size 100B.

send=ascii_set,recv=blind_read,conns=50,key_prefix=foobar,key_prealloc=0,pipelines=16,value_size=100 send=ascii_set,recv=blind_read,conns=50,key_prefix=foobar,key_prealloc=0,pipelines=16,value_size=100,thread=1 send=ascii_set,recv=blind_read,conns=50,key_prefix=foobar,key_prealloc=0,pipelines=16,value_size=100,thread=1 send=ascii_set,recv=blind_read,conns=50,key_prefix=foobar,key_prealloc=0,pipelines=16,value_size=100,thread=1 send=ascii_get,recv=blind_read,conns=50,pipelines=16,key_prefix=foobar,key_prealloc=1 send=ascii_get,recv=blind_read,conns=50,pipelines=16,key_prefix=foobar,key_prealloc=1,thread=1 send=ascii_get,recv=blind_read,conns=50,pipelines=16,key_prefix=foobar,key_prealloc=1,thread=1 send=ascii_get,recv=blind_read,conns=50,pipelines=16,key_prefix=foobar,key_prealloc=1,thread=1 send=ascii_get,recv=blind_read,conns=50,pipelines=16,key_prefix=foobar,key_prealloc=1,thread=1

I used the github memcached on an r4.4xlarge. I ran memcache-top on the server instance to measure the requests per second, showing about 750k gets/sec and 600k sets/sec.

With a ratio of 10:1 gets to sets I’m seeing about 3.5M req/sec which seems better than elasticache.

[dormando]

Please don't try to dial this back to win. I showed you how things work, go ahead and fiddle with them as you want.

1) sure, 100b, but that will just make it easier for the CPU version to hit the packet rate limit. I dialed it down to show just how fast the key rate is. Your entire proposal was that CPU bottlenecked the NIC, and it does not. Also, most people have 100b keys, nevermind the values.

2) 1:1 was never realistic. It's not even remotely realistic; as I said earlier 5:1 would be pessimistic. In reality the instances which have get rates in the millions tend to have 100:1 or better ratios due to the nature of the data they're caching.

Yes, the newer LRU algorithm doesn't grab LRU locks on the read path, so it'll scale with the number of CPU cores. As I said in earlier comments, the sets don't currently scale, especially if you're hammering the same LRU (which is again, unrealistic). If you just do a pure set load you'll land somewhere between 900k and 1.5m ops/sec.

3) I did both single-get-pipelined and packet-pipelined benchmarks; also absolutely not. Clients are designed to use the multiget mode when multiple keys are being fetched from the same server. This benefit is lost with the binary protocol (which will be fixed at some point).

4) Try an mget with 16, it won't be too far off, though you might have to add one more mc-crusher thread.

In your last test, you're simply overloading it with sets. If you want to mislead people with a test like this, go ahead; but I'll point it out.

3.5M/s isn't too bad.

Memcached really isn't a great target for your sort of work. I love the idea of FPGA offload, but trying to advertise your thing as superior by making up your own rules is going to get called out.

1) The popularity of redis is absolutely damning in general. if people are okay with the performance of a single CPU database with all-over-the-map latency profiles, the odds of you finding enough customers with extremely high rate memcached pools to sustain a business are essentially zero. You'd be solely tricking people who think they need it.

2) You are not facebook. Nobody is facebook but facebook. 100b is not representative. It's not even representative of facebook's load.

What's worse, even for a more common case, if 99% of requests are 100b, the average size of an item might be 8k. Which doesn't mean that there are a bunch around 8k, but there could be a few thousand items that are 50k-500k+ in size, getting hit 1% of the time, or even 0.1% of the time.

500x the bandwidth of a 100b request for the same processing overhead. It's almost always something they need: a request might fetch a couple hundred items from memcached, with just a couple of them being large.

This ends up making RAM be the greatest expense in the system. If so few users really need this performance, and the newer versions of memcached have a much higher perf ceiling, the extra features it has to drive down RAM usage are more valuable.

The best cost/power savings most users can do is find a way to get more RAM attached to fewer CPU cores: to be frank a r4.4xlarge would suit better with 8 cores. Or find ways push larger cold values into flash, freeing up RAM for those 100b values to be served quickly.

[dormando]

Try 12 workers to start, there are some bg threads.

I've plenty of experience with both hardware and virtual machines, I just don't use AWS myself much. I can get something like 800k read ops/sec from a 4core raspberry pi2, and I hope the AWS instance isn't that terrible.

with mc-crusher:

./mc-crusher conf/someconfigfile ipaddress port

https://github.com/memcached/mc-crusher/blob/master/conf/asc... - this is a decent read test with pipelining (give the test a few seconds to get through its sets). The inbound requests are pipelined, but it'll still send each get response in individual packets. This is what I use to test syscall/interrupt overhead.

https://github.com/memcached/mc-crusher/blob/master/conf/mge... this is the same thing, but with mgets. I'd copy the set line from ascii too:

send=ascii_set,recv=blind_read,conns=10,key_prefix=foobar,key_prealloc=0,pipelines=4,stop_after=200000,usleep=1000,value_size=10 send=ascii_mget,recv=blind_read,conns=50,mget_count=50,key_prefix=foobar,key_prealloc=1

can vary the value_size to and mget_count to see how that changes things. You can also pre-warm with the 'bench-warmer' script that comes with it, or remove stop_after and adjust usleep to adjust get/set ratios.

Watch top on the client host, and if mc-crusher is capping out its CPU cores, add more lines to the test but with the (confusing, sorry) threading enabled:

send=ascii_set,recv=blind_read,conns=10,key_prefix=foobar,key_prealloc=0,pipelines=4,stop_after=200000,usleep=1000,value_size=10 send=ascii_mget,recv=blind_read,conns=50,mget_count=50,key_prefix=foobar,key_prealloc=1 send=ascii_mget,recv=blind_read,conns=50,mget_count=50,key_prefix=foobar,key_prealloc=1,thread=1

That puts the first two tests on the "main" thread, then spawns an extra thread for the third test. you can keep copy/pasting that last line until the client or the server are saturated.

edit: sorry, the enc/compression question:

1) compression is typically done in the client to reduce bandwidth overhead. It's not very useful in the server.

2) encryption is becoming more popular, but doesn't currently exist much. The mainline OSS doesn't even have TLS support yet. Almost all use cases are on internal networks. FPGA's could potentially help there... aes-ni on intel cpu's isn't awful though.