Worker -> Durable Object fetch RPC call performance in Rust

Hello I'm running into some performance issues using Workers and Durable Objects. I'm using rust. I would like some help improving performance.

I have a worker that clients connect to. The Worker then passes the request through to a Durable Object (DO) which establishes a websocket connection and sends over a somewhat large batch of data.

I measured a measures for a cold start of the durable object.
* the total time for the Durable Object "fetch" method to complete

do_t

do_t

* total time for the worker to call the durable object fetch

worker_t

worker_t

* size of the payload the durable object sends

size

size

Then I take

Δt = worker_t - do_t

Δt = worker_t - do_t

to measure the latency for the worker to call the durable object fetch. I plot

Δt/size

Δt/size

to get the attached plot. The equation of the plotted linear regression is

y = 0.16x + 295

y = 0.16x + 295

so rate in increase in latency is

0.16 ms/KB

0.16 ms/KB

(or

50Mbs

50Mbs

) is this expected?. Also my interpretation of the

is the fixed time in

ms

ms

for my DO to boot up. Is there a way to make this faster? A

290ms

290ms

startup seems slow.

My main thoughts:
* Am I just doing something dumb to make this slow?
* How can I improve this scaling (the

0.16ms/KB

0.16ms/KB

)? Could it be caused by crossing a WASM <-> JavaScript boundary? If so, could I improve performance by rewriting this all in JavaScript.
* How can I improve the startup time?
* What are some things I should look for? I noticed I maybe be doing an extra allocation of the data, could things like that be significant.
* is there a better way to directly measure the Durable Object startup time than what I've done here?

I also attached some code snippets where I call the durable object from the worker, and the the

fetch

fetch

method:

foo.rs1.95KB

Worker -> Durable Object fetch RPC call performance in Rust

do_t

do_t

* total time for the worker to call the durable object fetch

worker_t

worker_t

* size of the payload the durable object sends

size

size

Then I take

Δt = worker_t - do_t

Δt = worker_t - do_t

to measure the latency for the worker to call the durable object fetch. I plot

Δt/size

Δt/size

to get the attached plot. The equation of the plotted linear regression is

y = 0.16x + 295

y = 0.16x + 295

so rate in increase in latency is

0.16 ms/KB

0.16 ms/KB

(or

50Mbs

50Mbs

) is this expected?. Also my interpretation of the

is the fixed time in

ms

ms

for my DO to boot up. Is there a way to make this faster? A

290ms

290ms

startup seems slow.

My main thoughts:
* Am I just doing something dumb to make this slow?
* How can I improve this scaling (the

0.16ms/KB

0.16ms/KB

fetch

fetch

method:

foo.rs1.95KB

Worker -> Durable Object fetch RPC call performance in Rust

Worker -> Durable Object fetch RPC call performance in Rust

Similar Threads

Similar Threads

Similar Threads