Scaling Memcached: 500,000+ Operations/Second(blogs.sun.com)
blogs.sun.com
Scaling Memcached: 500,000+ Operations/Second
http://blogs.sun.com/zoran/entry/scaling_memcached_500_000_ops
6 comments
Splinting memcache over several servers has all of the classic scaling issues. Including the need for 10Gig switches which are still stupid expensive:
Cisco 16-Port 10 Gigabit Ethernet Module Our Price $27,433.91
PS: You could probably get close with a 32 port GB switch with a 10GB unlink, but it's still not free.
Cisco 16-Port 10 Gigabit Ethernet Module Our Price $27,433.91
PS: You could probably get close with a 32 port GB switch with a 10GB unlink, but it's still not free.
A single threaded Redis instance on a 300$ Linux box performs 100,000 operations/second. Assuming a quadcore and running N instances to approach 500,000 op/seconds in cheap hardware should be not hard, so I'm not very impressed by this numbers, especially given the hardware used.
Read the facebook engineering notes on memcached. Linux kernel needs to be patched to fully utilize all cores even at 200k qps with 173ms average latency. Also you didn't give average request size and latency numbers, which makes your number less helpful.
Seconding this.
Besides CPUs & RAM constantly getting faster, since when was handling in-memory data fast in any way impressive?
Besides CPUs & RAM constantly getting faster, since when was handling in-memory data fast in any way impressive?
It does advertise Sun's hardware but scaling memcached to heavily multicore setups is interesting, irregardless of hardware used - same points would be valid on highend intel servers. Good to know it's possible to saturate 10G without much tweaking.
A typical "drag race" benchmark. There is no latency numbers for the corresponding throughput. The facebook engineering notes was a little more helpful, with at least average latency 173ms at 200k qps with an intel 8-core box with 1GE NIC (they could push it to 300k but the latency was too high.) So a max 500k qps (with unknown latency) on a much more expensive T2 with 10GE NIC is not that impressive.
I'd like to see median and 99 percentile latency for any throughput numbers.
I'd like to see median and 99 percentile latency for any throughput numbers.
also see the older http://www.facebook.com/note.php?note_id=39391378919
is this an ad? serious question.
http://www.google.com/products?client=safari&rls=en-us&#...
Take a look on Newegg, and it would seem that it'd be enough to buy you around 20 Intel boxes with 16 gigs of ram each. The configuration I used put a Core2 Quad on each machine, so our cluster would have 80 cores and 320gb of memory. Unless there's a huge gotcha somewhere regarding the system bus in the Intel architecture vs. the Sun architecture, the Intel cluster still seems to win out. The Sun box indicated that they used DDR2 chips but didn't indicate what speed, so I picked DDR2 800.