Topic: Testing Armory's encryption parameters, results and questions (Read 1439 times)
Armory's KDF is a precursor of Scrypt. It is mostly memory bound, where the KDF keeps hashing pages upon pages of data that it has to keep in RAM because the final result requires data from unpredictable page indexes. What the user gets to control is the amount of RAM used for the most part. This makes stuff like GPU brute force attacks inefficient.
Bumping this because I also had this question and I found the answer to it.
The target compute time is user set and 250 ms by default, Armory uses a KDF with SHA512, and a salt generated somewhere, to implement the intensity in the wallet encryption, and the KDF's parameters are calculated here. The compute time influences the number of iterations, and the amount of memory that it'll use, which is itself limited by the memory limit that you set (by default 32MB).
Inside the KDF parameter calculation code I linked, there is a loop that looks for the optimal amount of KDF blocks (and hence memory) to use for the KDF. This is doubled each time in the loop, which roughly doubles the amount of time KDF takes:
Eventually there comes a breaking point where increasing the target compute time doesn't make the KDF take much longer because the memory limit in place will limit the amount of memory KDF can use, making the encryption process more memory-bound than time-bound.
1. What determines the target compute time?
(examples: number of AES rounds, key stretching derivations, etc.)
(examples: number of AES rounds, key stretching derivations, etc.)
The target compute time is user set and 250 ms by default, Armory uses a KDF with SHA512, and a salt generated somewhere, to implement the intensity in the wallet encryption, and the KDF's parameters are calculated here. The compute time influences the number of iterations, and the amount of memory that it'll use, which is itself limited by the memory limit that you set (by default 32MB).
2. From my limited testing, it seems like raising the target compute time increases the required unlock memory. In fact, I went and did more doublings of target compute times, and it seems like for every doubling in compute time, the unlock memory doubles as well. What is the reasoning behind this?
Inside the KDF parameter calculation code I linked, there is a loop that looks for the optimal amount of KDF blocks (and hence memory) to use for the KDF. This is doubled each time in the loop, which roughly doubles the amount of time KDF takes:
Code:
// Start the search for a memory value at 1kB
memoryReqtBytes_ = 1024;
double approxSec = 0;
while(approxSec <= targetComputeSec/4 && memoryReqtBytes_ < maxMemReqts)
{
memoryReqtBytes_ *= 2;
sequenceCount_ = memoryReqtBytes_ / hashOutputBytes_;
lookupTable_.resize(memoryReqtBytes_);
TIMER_RESTART("KDF_Mem_Search");
testKey = DeriveKey_OneIter(testKey);
TIMER_STOP("KDF_Mem_Search");
approxSec = TIMER_READ_SEC("KDF_Mem_Search");
}
memoryReqtBytes_ = 1024;
double approxSec = 0;
while(approxSec <= targetComputeSec/4 && memoryReqtBytes_ < maxMemReqts)
{
memoryReqtBytes_ *= 2;
sequenceCount_ = memoryReqtBytes_ / hashOutputBytes_;
lookupTable_.resize(memoryReqtBytes_);
TIMER_RESTART("KDF_Mem_Search");
testKey = DeriveKey_OneIter(testKey);
TIMER_STOP("KDF_Mem_Search");
approxSec = TIMER_READ_SEC("KDF_Mem_Search");
}
Eventually there comes a breaking point where increasing the target compute time doesn't make the KDF take much longer because the memory limit in place will limit the amount of memory KDF can use, making the encryption process more memory-bound than time-bound.
I'm playing around with Armory's wallet creation, and noticed two parameters that the user can determine: "Target compute time" and "Max memory usage". Question:
1. What determines the target compute time?
(examples: number of AES rounds, key stretching derivations, etc.)
To test the wallet's encryption properties, I created a few test cases with the following parameters and recorded the following results:
Target Compute Time 250 ms
Max Memory Usage 32 MB
Password Length 32 64 128
Time to Unlock 249 ms 246 ms 246 ms
Unlock Memory 4 MB 4 MB 4 MB
Target Compute Time 250 ms
Max Memory Usage 64 MB
Password Length 32 64 128
Time to Unlock 164 ms 255 ms 165 ms
Unlock Memory 4 MB 4 MB 4 MB
Target Compute Time 500 ms
Max Memory Usage 32 MB
Password Length 32 64 128
Time to Unlock 332 ms 334 ms 501 ms
Unlock Memory 8 MB 8 MB 8 MB
Target Compute Time 500 ms
Max Memory Usage 64 MB
Password Length 32 64 128
Time to Unlock 336 ms 501 ms 498 ms
Unlock Memory 8 MB 8 MB 8 MB
2. From my limited testing, it seems like raising the target compute time increases the required unlock memory. In fact, I went and did more doublings of target compute times, and it seems like for every doubling in compute time, the unlock memory doubles as well. What is the reasoning behind this?
1. What determines the target compute time?
(examples: number of AES rounds, key stretching derivations, etc.)
To test the wallet's encryption properties, I created a few test cases with the following parameters and recorded the following results:
Target Compute Time 250 ms
Max Memory Usage 32 MB
Password Length 32 64 128
Time to Unlock 249 ms 246 ms 246 ms
Unlock Memory 4 MB 4 MB 4 MB
Target Compute Time 250 ms
Max Memory Usage 64 MB
Password Length 32 64 128
Time to Unlock 164 ms 255 ms 165 ms
Unlock Memory 4 MB 4 MB 4 MB
Target Compute Time 500 ms
Max Memory Usage 32 MB
Password Length 32 64 128
Time to Unlock 332 ms 334 ms 501 ms
Unlock Memory 8 MB 8 MB 8 MB
Target Compute Time 500 ms
Max Memory Usage 64 MB
Password Length 32 64 128
Time to Unlock 336 ms 501 ms 498 ms
Unlock Memory 8 MB 8 MB 8 MB
2. From my limited testing, it seems like raising the target compute time increases the required unlock memory. In fact, I went and did more doublings of target compute times, and it seems like for every doubling in compute time, the unlock memory doubles as well. What is the reasoning behind this?
Jump to: